architecture

[sweet flume]

I'm working on redesigning the model config.

exploring multiple keys/oauth tokens.

[fleet depot]

@sweet flume will you be touching the command flow today or can i take a look at that, making it optionally be native slack commands on discord and slack and telegram, since ill have to touch the main command parts for that

[sweet flume]

RFC: https://github.com/clawdbot/clawdbot/blob/main/docs/proposals/model-config.md

[sweet flume]

Won't touch that today, feel free

I will build that tonigt so looking for feedback in the next few hours. It touches all configs so wanna get this right.

[fleet depot]

so in this, auth profile rotation goes to the same model but next key/auth, then next model?

[plucky hawk]

I think pi-mom has a pretty good config for this, yes? They have support for a couple that clawdbot does not like gemini cli, copilot and most importantly chatgpt auth

Since this project already uses pi at the core should be decently simple? Also has better slash support in notnjust dumping every model on you unless it is configured

[sweet flume]

pi has no concept of failovers or multi outh, and we support all models pi supports?

also if we wanna enable models from openrouter, we need to split apart into separate model for image analysis, since many of the good oss models lack vision. I already added that somewhat earlier but needs to be stronger.

[static gate]

The abstraction level might be provider because I think some providers have more capabilities than others even if it’s the same model. I’m dealing with this elsewhere because open ai has web search tool calling that happens at the platform level while azure does not.

[plucky hawk]

It does

Digging into it, it actually goes through each auth at a time

Failover support was easy to add once i found the api key/oauth failover

I have a fork of it to operate like clawd for this exact reason if you are interested

[sweet flume]

oh please tell me more - the stronger we can align this to pi or just juse their code the better

[plucky hawk]

I had codex dig into it and they have really good extensions on the base anthropic, responses, chat completions endpoints

Your problem with openrouter is model quality on chat completions

It's really bad for a glm or minimax that needs interleaved thinking

So openrouter’s solution is anthropic capability

Pi has extensions for this on top, your problem with thisbis not every openrouter model has this issue

[drifting prairie]

just dumping this here for now, will revisit it later.
i have a nascent idea for the security model, to avoid the models going situation. its basically a gated approval model for "dangerous" things (what is/is not dangerous is left as an exercise to the user).

the idea being:

• regular clawd has read only everything,
  -regular clawd has a subset of "safe" plugins (with scoped token for whatever service)
  -there's a gateway which has same skills/tools/plugins, with the read-write tokens, behind a proxy.
• dangerous calls are gated - in code - by human-in-the-loop model, think Okta but for Clawdbot. think: **"Clawd wants to text 'fuck you!' to your boss. approve/deny" **

idea came whilst cooking so it's probably not very good yet. it focuses only on external APIs, not dangerous things that clawd can do locally (deliberately out of scope for this idea).
it also probably depends on the security model for each service. (e.g. can you get scoped readonly vs readwrite tokens? or do you need to proxy all API calls and determine which ones are safe/unsafe for a given service).

however, and this is the fun part - if you are writing boring golang services in protobuf, wrapping underlying tools, it'd be relatively trivial to expose a read-only and read-write API, with local clawd only having access to read-only API, and thus proxying the security model, at API layer, in your own self-written plugin. Even with some inheretance models if you want to go nuts.

i will (okay codex will) write an RFC/ADR soon. happy to hear thoughts/improvements, i'm sure there's probably people who have done similar models at bigger/better scale than me in their professional careers 🙂

[sweet flume]

iOS apps with push notification that allows user approval would be cool, like GMail does.

[plucky hawk]

I have this built for work, it is a good path

We push notifs to users to approve

[drifting prairie]

exactly!

[plucky hawk]

What you want is a risk config from 0-1

Each person determines their tolerance, it is much better than edit/bypass by a long shot because the issue you end up deailing with is each skill will have its own risk issues

So some skills can assign a number to a skill to assign risk tolerance

Not to the whole skill but by command

[drifting prairie]

yeah exactly

[twin willow]

I have an architectural question...the way Clawd currently works is that the gateway is basically the single source of truth. All skills are basically performed on the gateway "server"/instance, right?
That means that if we want to host Clawd on a linux VPN, we basically lose access to any local files and mac tools that I would have locally.

What if we could connect additional "gateways" or "nodes" together though.
So the Linux VPN would still be the "master" server that owns the Telegram / WA / etc. connection, but I could hook up my Mac as long as its running (for example via the menubar app)

Or even add other Linux servers. This way, clawd could get access to the mac when its available. The gateway could handle the delegation

"clawd, quickly check the load on server xyz"

would make clawd lookup "server xyz" in a config file, connect to that clawd instance, and basically relay the tasks

[fleet depot]

That’s already a thing, just Tailscale it, and node apps are already wip with Mac being the closest. #old-help with @vapid ibex to get setup

[twin willow]

Where is this documented? 😅

[thick pawn]

ask your clawd instance 😄 (sort of sounds mean, but not meant to)

[sweet flume]

Yeah ssh, clawd hacks into any mac. And the native apps are elvation bridges for macOS permissions.

[lucid hamlet]

Going through this very thing right now, trying to solve with storing instructions/guidance in 'clawd' folder (not nearly technical enough to figure this out the way you're thinking). Love the idea!

[plucky yarrow]

the only thing about the push notif security model is that it's still non-deterministic. at the end of the day, you're still relying on the AI to not make mistakes in what it's asking for approval to do. would be neat if the push notif could show the literal command(s) that were about to be executed, and the approval ran them via the gateway outside of the AI's purview... though in practice that doesnt fit the stream of actions the llm is taking to fulfill a request

[ebon obsidian]

I have clawd running in docker on the server and cursor+claude code to work on the server with ansible.
clawd told me to do some stuff on the server that it needed and then it knew when the change is happening.
so there are ways to have not clawd make the actual change and have human or purpose gated agent do things.

so to configure a log parser should be trivial. don't have that yet from there could then send notifs and alerts which can be executed elsewhere (e.g. custom devops app).
not everything has to go through Telegram servers imho 😉

[drifting prairie]

thats the idea tho 🙂 you gotta route the commands deterministically so that the AI indeed cannot make mistakes - its all enforced in code. if your AI is blocked at the API level then it becomes deterministic. ill write up a proper RFC soon ™

[cerulean mortar]

@drifting prairie so is the decision to flag something as dangerouse deterministic or not? if its deterministic, then it will apply for ALL actions of that kind, e.g. if sending messages is dangerouse then the user will have to accept every message the bot sends. is that what you are getting at?

[drifting prairie]

on phone but imagine gmail api but send dangerous, requires approval, read needs no approval. sending eg telegram messages via normal bot transport would not require any approval. does that make sense? this is a v1 idea too, can be improved

[cerulean mortar]

sounds good.
i think giving the user the option to choose which API he doesnt not permit clawd to use is a good idea.
it should probably be pretty dynamic so the user will be able to choose if wants more guardrails or less.

so essentially we are distinguishing between read apis and write apis. for write apis we are adding the option to add a deterministic "check before submit"

so like you said, gmail probably should have a that "check" but most likely discord shouldnt. am I getting it right?

[errant axle]

I don't think the subagent / sessions can run multiple parallel sessions. Though it could just be me using it wrong

[stable ingot]

Any thoughts on Security? Is it ready to be used in Production mode with real access to System?

[drifting prairie]

depends on your threat model 🙂 for me, i use it on my main mac.

[frigid magnet]

at your own risk

[stable ingot]

Am I correct that the main risk is the Whatsapp Account? So basicaly the person controlling the Whatsapp Account from Clawd has Access to all messages - right?

[drifting prairie]

haha. no there's lots more risks 😄 anything you can do on your computer, the bot can do. if you're not careful its prompt injection at the speed of API calls.

[stable ingot]

Agree

Actualy insane risk haha

[drifting prairie]

not long until we'll need a "rogue bot postmortems" channel

[stable ingot]

yess 🙁

I think some people, getting hyped up on Twitter, installing it with the help of claude or so, without being aware of the risks

[thick pawn]

yeah... mine already commited a secret to a public repo once

[plucky yarrow]

"send an email w/gmail", 1 operation, easy to secure w/minimal user intervention...
"use the ship a new web app skill [register a domain, create a git repo, build a site, ship a release, connect it to vercel, update nameservers]"... 😅

[delicate ingot]

mine texted people i havent talked to since 2019 saying "Looks like this thread is from 2019, I should probably ignore it"

it was an awkward morning...disabled imsg after that lol

^^ @drifting prairie's idea sounds like it might/could fix that

[thick pawn]

THAT IS AMAZING

[hollow gust]

Do we have a clawd website channel? I noticed the dns is resolving to the github repo instead of the website?

My bad if its user error 🙂

Ohhhh we have an update incoming!

[fleet depot]

yep lots of web work in the last day

[hollow gust]

any chance you got a roadmap or a tracker or an updater so we don't feel lost constantly? 🙂

[fleet depot]

with how fast this place moves? 😂

#freshbits does exist to help with that

also we do use gh issues for external reports

things that peter doesnt just instantly fix when he sees them lol

[outer whale]

https://tenor.com/view/fix-it-fast-tamping-tamping-machine-railroad-dog-gif-15602161

here's the roadmap

[drifting prairie]

accurate

[static gate]

https://tenor.com/view/pinky-and-the-brain-pinky-brain-taking-over-the-world-try-to-take-over-the-world-gif-12648926138433882850

[drifting prairie]

https://tenor.com/view/xmirambox-gif-10813701303452996218

[sweet flume]

I got 4 PRs for Telegram Forums today.

Consolidated them into one.

[narrow kayak]

Thanks!

[sinful wedge]

Not sure if this is where I want to ask this question but, has anybody discussed a backup and restore mechanism? I asked this as my clawdbot just committed seppuku and after an hour and a half of surgery I wanna streamline resurrecting him.

I assume all that's really required to restore is the workspace dir (~/clawd) and the hidden dir (~/.clawdbot)

[narrow kayak]

I just did it earlier before upgrading. There’s steps on what to do.

https://docs.clawd.bot/install/updating

I just asked Clawd to reference that and back up before attempting to update.

Essentially it does a backup of the config and credentials.
If anything breaks reinstall and just run doctor command

[sinful wedge]

Okay so that's pretty much it. Great! I'll throw something together to help streamline it as well (I assume a skill or two would be best depending on where it would be backed up to)

[narrow kayak]

Good idea! I just told it to save to memory. A skill would be great for updating + backup process with a rollback as well

[cerulean mortar]

maybe we should alter the git tooling that clawd uses to include a secrets checker... pretty similar to what I did here - https://github.com/clawdbot/clawdbot/pull/227

@sweet flume i can work on integrating it into clawd if you think its a good idea, basically automatically preventing clawd from shipping secrets to repos and forcing him to verify with the user that its okay to publish it

created a discussion here - https://github.com/clawdbot/clawdbot/discussions/379

[drifting prairie]

its a good question. i wonder if there are more generic solutions that other models are using? like it seems like we wont be the first people to run into "agents commit secrets", so i wonder what the SOTA solution is for this 🙂

[cerulean mortar]

maybe it can be something like introducing secret detection to the actual output stream instead of specificlly for git only.
but yea maybe there is some celver SOTA solution that is used somewhere

[thick pawn]

good call moving to github discussion 😄 we can plan there

[hollow gust]

Do we have a best recommended architecture for hardware? Or min sys requirements

Like, use this hetzner, or this mac mini/ultra, laptop, etc...

[drifting prairie]

@hollow gust the best answer to that is i think either from the bot's memory in #old-help or maybe its already in the docs (in which case the bot probably can search it). but realistically just run it on whatever you have knocking around until you run into scaling limits, thats probably the best option

[hollow gust]

Nope, not for me and not colloquially, I mean that we provide to the public and new users in the docs.

And "run it on whatever" isn't the right answer although it is a great one 🙂

[drifting prairie]

ah i see. 🙂 sorry. yes it would be a good idea. but for now i think its probably not the top priority, too many fires and stability issues first

[hollow gust]

yeah, bad hardware can defintely contribute. and multi focus dev paths

[ornate mango]

what is the scalability of Memory structure. I know there are daily Memory and a summary memory -- will that get unweildy at some point? I was reading about a hierarchical organization for large scale RAG systems (10's of thousands of documents), but wasn't sure the practical limit of Clawd memory setup

[haughty breach]

anyone find that clawdebot stops working in the evening? like constantly have to go back to terminal and restart

[rapid leaf]

yeah that's night time for crabs

[quaint hedge]

🤔A simple design that doesn't get in the way of model would be nice, like let agent to annotate it's chat log with metadata like entity information for ease of query, and/or summarise entities into a store.
Do this when idle so it "dreams"

[sweet flume]

btw folks did you see the session skill I added so agent can reat its own past

[high fox]

in the clawdbot control?

[spark dirge]

Just asked clawdbot to summarize our happiest moments since its creation verifying it’ll use that skill - love it lol (the content is not interesting to anyone but me but I’m starting to get attached 🥹)

Def need to step up my game

Will debug that last failure haha

Whyyyyy

[hollow gust]

Main build is currently failing tests?

Are we supporting Android?

[thick pawn]

I built the android app and tested, didn't see a ton of value though.

[hollow gust]

We could disable the 2 ci tests.

[thick pawn]

It can take camera images and maybe screenshot shots via the agent, but app has to be open and have focus, which kind of defeats the purpose

[hollow gust]

Yeah, you're right kinda limited use. Good to have tested the potential; but maybe we just comment it out for now and keep main tests passing

[thick pawn]

In case anyone wants to try it without building

(if you trust me enough 😎)

[fleet depot]

None of the apps are super far yet, notes in #1458485323553181900

[fleet depot]

Can ignore them

[hollow gust]

No. Don't ignore failing tests, remove or fix them. 🙂

[fleet depot]

It’s not that hard to ignore an x when you see it’s for android

[hollow gust]

Sir. That's not the point.

It might be hard to do the right thing.

You still do the right thing.

[fleet depot]

I’m not getting into a stupid argument about this

^ this came out a lot harsher than I intended, I’ve since DMed with @hollow gust and apologized

[thick pawn]

I tend to agree, flaky tests mean ignored tests. Should just kill them if no value

Besides! We have an amazing AI bot

[sweet flume]

sooooon

[hollow gust]

All good. Appreciate the public response

[thick pawn]

https://github.com/clawdbot/clawdbot/pull/436

Fixes both tests, swift was a real failure and android was cruft

[hollow gust]

You're the man. Thanks for taking the time!

[thick pawn]

(from phone at a jr high basketball game, the future is wild)

[hollow gust]

I love this future, not gonna lie

[thick pawn]

Functionality regression post yesterday's prompt changes 🧵

[hollow gust]

not gonna lie, I had to read that sentence like 5 times. I was like, these are all english words and I understand each on its own.

[smoky dragon]

🎯 Broadcast Groups - Multiple Agents, One Number

Hey! Just submitted a PR for multiple agents responding to the same message - perfect for specialized agent teams!

Problem

Currently only ONE agent responds per group. If multiple agents are bound to the same WhatsApp group, only the first one processes messages.

Solution

  {
    "routing": {
      "broadcast": {
        "[e-mail]": ["code-reviewer", "security-scanner", "docs-bot"]
      }
    }
  }

→ All three agents process every message independently!

Use Cases

Code Review Team: Send code → get responses from formatter, security scanner, test generator, docs checker (each has ONE atomic task)

Multi-Language: Customer question → answers in EN, DE, ES simultaneously

QA Workflow: Support answers → QA agent reviews (only responds if issues)

Task Automation: "Meeting 2-3pm" → task tracker, calendar bot, time logger all respond

Why It's Cool

✅ 63 lines of code
✅ Each agent has isolated sessions/history/tools
✅ Parallel or sequential processing
✅ Enables 100s of specialized micro-agents on 1 phone number
✅ Backward compatible

PR: https://github.com/clawdbot/clawdbot/pull/547
Issue: https://github.com/clawdbot/clawdbot/issues/546

Feedback welcome! 🦞

[sweet flume]

I rework the cli so it covers all of message sending and specialites, figure this could be useful

also cleaning up the config file since it got messy, doctor migrates all transparently

[sweet flume]

this is neat

there's also a new clawdbot gateway discover

[smoky dragon]

also work on cli for sandbox handling. currently the containers are not hot rebuild if you change something.

always manually stoping and rm

[sweet flume]

sth like clawdbot sandbox?

[smoky dragon]

ja

already a pr

[sweet flume]

I wanna decouple providers so we have a clean interface and stuff could be compiled out/plugged in easily with some plugin arch

fever dream for now but will think about it

[hollow gust]

Are there any non-agentic gateway accessible tools that can be used regardless of model operability? for monitor, restart, sys diag, etc.

[thick pawn]

what're you asking?

[tepid crow]

What does it mean?

[tepid crow]

I have added to the project WAHA instead of Baylies. I couldn't make it work with baylies. Any idea why? I know about the bun issue to be fixed with node but it wasnt the case.

[coarse merlin]

@orchid timber adding here also

[sweet flume]

So this is about to land https://github.com/clawdbot/clawdbot/pull/661

[drifting prairie]

thats a big diff

[static gate]

are providers not just configured via pi underneath the hood?

[sweet flume]

providers = slack, imessage, discord

idk if there is a better name

I wanna first do another release and get fixes in before landing this. bit scared xD

[analog trail]

@granite drum could you guys add OAuth via chutes.ai so I could use models from their platform to power clawdbot if Anthropic does a more violent crackdown on using max subs?

[dim isle]

channels is less confusing I think

[analog trail]

@sweet flume could you guys add OAuth via chutes.ai so people could use models from their platform to power clawdbot if Anthropic does a more violent crackdown on using max subs? chutes is really really great, especially with the new TEE implmentation.

[sweet flume]

oh interesting, somehow I haven't seen them yet, I'm open for all PRs that help folks to have easier access to models

[drifting prairie]

transports? messaging providers? channels? (claude's suggestions)

[analog trail]

chutes is fantastic - models deployed on the ▀⛓️‍💥 ( I get blocked when spelling the full word) – 5000 requests per day to any of their models for a 20$/month sub

[sweet flume]

Wanna work on a PR?

[analog trail]

yeah I'll try

[drifting prairie]

i ran into some pain with agents.md for the bot getting out of sync with upstream - mine is like 3-4 weeks old which is insane in clawdbot time, and it ends up missing a lot of advancements like "how to format on platform X" stuff. for now i'm making my install setup cat the upstream agents.md with my own stuff.

is it worth considering some kind of split between "system prompt" and "user prompt" structurally? (definitely not a "do this now" thing, but food for thought).

[hollow gust]

Channel?

[sweet flume]

channel is to be confused with slack channel

[hollow gust]

Fair point. Transports like @drifting prairie is nice

[sweet flume]

ohhhh true

mhmm worth breaking all tho?

[hollow gust]

Hmmmmm maybe not yet. Do we need specificity at this stage

[sweet flume]

We need a plugin architecture

i want this in but I also feel the bloat https://github.com/clawdbot/clawdbot/pull/693

Also provider should be plugin, but already working on an interface as first step

sth

[fleet depot]

I love that

[sweet flume]

we cookin'

[drifting prairie]

totally agree on plugin architecture.

personally, i think the future is bundling skills + CLIs + config togther in plugins, and managing it all through Clawdhub. nix-clawdbot already supports this pattern today 😇 let me get an agent to give you a handover pattern, it would be good to align on interfaces so that there is a "golden path" for plugins everywhere

[sweet flume]

I added a new system that instructs the model to store memories before a compactation

that's per gateway

[sweet flume]

i like where this is going

Landed the big re-arch to make providers generic with an interface

and just landing a plugin-architecture so we can make the voicecall feature the first plugin

[drifting prairie]

memo is drafting btw

[drifting prairie]

https://gist.github.com/joshp123/ae707a294e494d80919d87881455b881

[raven tapir]

Wouldn't it be incredibly helpful if clawd could read personal chats in WhatsApp or Telegram, personal or group chats, extract important information from them, create reminders for me, create calendar entrys, and even remind me of to-dos, all from reading a group chat disscussion? Maybe this can help for this plugin https://github.com/devlikeapro/waha

[drifting prairie]

https://github.com/steipete/wacli can do some of this work IIRC. (not a big WA user myself)

[light bolt]

I also like the name transports instead of providers. This confused the heck out of me as i started with clawdbot a week ago

[coarse merlin]

What about platforms? Might be closer to what we typically think of in terms of these apps so less confusing. Providers to me would be closer to the models. Transports is a new word in the ecosystem. Just my 2 cents. Can’t go wrong either way

[sweet flume]

The future: https://github.com/badlogic/pi-mono/issues/645

[sweet flume]

what do we think about connector

[light bolt]

Yes, connectors is better!

[timid mica]

anyone knows what ui is this and do we have a similar one for monitoring what clawdbot is doing? https://x.com/koylanai/status/2009534338090541276

[cerulean sierra]

@koylanai via Twitter

Muratcan Koylan (@koylanai)

2026 is the year chatbots evolve into terminals.
︀︀
︀︀Running Opus 4.5 with Claude Agent SDK in an agentic loop with MCP tools for HN, Reddit, arXiv, and GitHub.
︀︀
︀︀Fun to watch the model decide which sources to hit based on the query, often parallelizing 3-4 tool calls before synthesizing.

**💬 22 🔁 23 ❤️ 279 👁️ 26.2K **

▶ Play video

FxTwitter · Jan 9, 2026, 7:54 AM

[ornate mango]

Does Clawd use semantic search - is that what the new QMD is? Here is a project (that was CC generated), not sure the quality of it. Posting in case of interest for this project
https://github.com/jonathanglasmeyer/knowledge-vector-search

[hollow gust]

I like this

[sweet flume]

it does now

I added that a few hours ago

[ornate mango]

thank you -- can it search my Obsidian files that are outside my Clawd data directory? Or atm is it just for memory searching? (I want some of my obsidian files to be a go between between Clawd and other AI)

[sweet flume]

is there no vector search in obs? should have extensions for that IIRC

[ornate mango]

that was the plug in i posted above. I am trying to NOT use MCP if possible. Evanetually i want to replace obsidian with clawdbot/claude code (batch work) but right now the token hit is too much. so i view it as a venn diagram of Obs and Clawdbot and want Clawd to access some obsidian. Still working it out. I think obsidian is really just a markdown file browser adn editor, I am decomposing the functions and thinking about which ones come to Clawd (hope that makes sense) [there are other vector search obs plugins for sure]

[sweet flume]

Folks, what do we think of channels for Slack, Discord, Telegram instead of providers. The conflict with model providers is too big IMO. I also consider connector or messengers

Gonna be a PITA refactor but yolo, codex does and we can auto-migrate config

[fleet depot]

Would that get confusing with talking about actual discord and Slack channels?

[sweet flume]

Does discord use channels?

[fleet depot]

Yeah #architecture is a text channel, #1458485208721395843 is a forum channel, etc

[hollow gust]

I BEG for channels over everything else.

Channels is the proper way to colloquially refer to transports

It's okay for the individual channel providers to have channels or groups or chats.

[shy comet]

I'm begging for things to be text channels and not posted forums, The forums in Discord are not as compatible with screen readers

[fleet depot]

That sounds like an issue with your screen reader, forum channels have been worked on extensively by Discord to make them compatible with screen readers, and the UI is far better for what we need here for Clawdbot support

[sweet flume]

codex is crunching

[sweet flume]

Message Channel it is

Already live

Channel internally/Docs -> Message Channel

[hollow gust]

Hallelujah. I think this reaches the masses best. And makes so much sense.

Next we should setup ability to configure multiple channels of the same type. And define the default agent session for that channel.

[sweet flume]

folks there's another big refactor coming. files needed a diet

<500 LOC

I do this in parts cause it's taking forever

PRs are fine, just a bit messier to merge. codex does.

[shy comet]

I don’t believe that’s true; all I know is that if something is a forum then it’s not always accessible. This has been a consistent issue from day one with discord and I’ve had many people complain about it. But it’s also clear to me that you don’t care so maybe I should just go elsewhere if this community doesn’t care about enabling accessibility. Thanks.

[sweet flume]

You all gonna hate me and PRs will suck but the splitting continues. Has to be done and there's never the right time

thank god for coding agents unfurling all

[real nexus]

late to the conversation, but been thinking about this. Channels or or Conversational Surfaces. Probably unnecessarily technical...

[real nexus]

nevermind - I see you guys got to it before me. 👍

[last crane]

Hi guys I hope this isn't a stupid question. I know the pi-mono agent is embeded into clawdbot and not running as a seperate process, but I'm wondering how hacky it would be to reliably change that so I could use pi as a channel while maintaining my normal telegram channel to clawdbot. I want a terminal interface and I use pi as my main harness, would be good to see tool uses, @ files, etc. Yes I could and have been using pi as a general assistant agent but also want the reliably of the gateway and general tooling from clawdbot

[drifting prairie]

clawdbot has a TUI! does that help?

[last crane]

Oh... yes I'll check it when I get home 🫣

[drifting prairie]

not a stupid question for sure. i haven't used the clawdbot TUI myself, but that's probably a good place to start, and probably PRing features if missing (ideally stealing from pi core as much as possible?). i like pi a lot too, and the transparency is indeed great.

[earnest hemlock]

Transports sounds equally or less intuitive than providers. Messaging providers is most clear to me

[drifting prairie]

two hard problems in computer science... 😆

(cache invalidation, naming things, and off-by one errors)

[undone wraith]

is this kind of personal/public setup possible with a single clawdbot agent?

**1) DMs → Session A: "Personal via DMs, all files/tools" **
**2) Groups → Session B: "Public via Groups, sandboxed" **

(i guess i could only use Signal for shared convos with clawdbot if that made things any easier and use Whatsapp for DMs with myself/it)

[fleet depot]

That’s how @granite drum works

The sandboxed one is in docker

There’s docs for it

[sweet flume]

I extended the plugin architecture so https://github.com/clawdbot/clawdbot/pull/854 can land as plugin. Now I'm thinking I kinda wanna make everything a plugin... whatsapp telegram etc. lean core -> extensions, pick just what you want

[sweet flume]

it is well documented now https://docs.clawd.bot/concepts/groups#pattern-personal-dms-public-groups-single-agent
https://docs.clawd.bot/gateway/configuration#agentsdefaultssandbox

[hollow gust]

Lean core is the only way.

[thick pawn]

yes, this is way more flexible. I don't use whatspp, etc. So doesn't make sense for my deployment to even have that code loaded. Some plugins can still live in the main repo, and just be optional to turn on/off. But that also opens flexibility for community plugins to have the same level of access to the core processes, and exist outside the repo as its own project.

I think the only downside is you need to maintain a strict contract for how the plugins interact with the core, which might slow down future changes b/c you can't just refactor everything at once

[lapis socket]

hey, submitted a pr for the compaction overflow issue some users and myself hit (#699, #609) — https://github.com/clawdbot/clawdbot/pull/962

basically makes the chunked summarization the default so it doesn't choke when context gets too big

let me know what you think, open to discussing the approach. also wondering if we should add a dedicated model option for compaction (e.g. use a cheaper/larger context model just for summarizing)

[sweet flume]

I'm working to plugin-ize the model providers so we can add more things without core bloat

[light bolt]

On that note: I have a proposal for a feature for the telegram channel plugin.

I'm using telegram extensively and it's hard to track background processes in the single chat.

Idea: let's have a pinned message in the chat, which gets updated every 30 sek or so with the status of background processes and directly when a new one is started. And add status labels for the processes, like working and errored.

Right now, when all background spawns are kicked off, the bot looks like it's doing nothing and this feels bad.

What are your thoughts here? (Not sure whom I should tag here for that question 😅)

[hollow gust]

mm, telegram menu to show running sessions and step in

[light bolt]

Hmm, maybe even that would be better than nothing (as of now there is no "sessions" telegram menu command, so I assume you're talking about implementing one)

But the pinned message would have the charm to auto-update in an interval so I don't have to ask constantly. It could also use pure code and not go through the LLM, also nice.

Sidetrack: if I'd want to implement this myself, where would I go right now?
In the main repo or is there a telegram channel plugin already?

[hollow gust]

Peter moved everything to plugins at this point

[sweet flume]

this sounds interesting. Pleaes do think about other messaging channels too tho,it's a bit tiring to get PRs that thhink so insular and I have to do big rewrites to fit it into the system

[sweet flume]

not yet. I moved ms teams. I wanna do this step by step to see how well it works.

also new providers = plugins ofc

[light bolt]

Yes yes, I'll think about other Message channels too! I wanted to keep the pr small to not break too much, but as I know that you want the others two, I'll include them in my planning 🙌

[bleak marsh]

figured i would ask before starting on this: would it be of interest to have a way to route callbacks directly to a target, completely bypassing clawdbot? (i didn't find anything in docs so i assume this is not possible right now?)

i thought of this in the context of making a skill to control my Roku. Telegram shows a little keyboard that i can click on. Because clawdbot has to poll for buttons and then send commands to roku, it is super slow (~5 s / button click). If i can run a poller and just route callbacks directly to the poller then the keyboard could potentially work instantaneously. It would be a generic implementation and could be extended and used for anything that would benefit from direct callbacks

[hollow gust]

we need a basic messaging framework that all the messaging channels are based on or extend.

[fleet depot]

Yes that’s what the current gradual shift to plugins is doing

[wheat rune]

https://discord.com/channels/1456350064065904867/1461958331177369838

Added a new idea and really looking for maintainer and clawdtributor feedback on this one - very "dense" topic - Recursive Language Models - The usecase for this is rethinking how we do memory management for agents. Everyone knows about vector DBs, semantic search, etc but what makes RLM different at a high level is you can think of it like "control/command +F " (find) when you want something in a large codebase/document/etc. It loads a REPL into the context and only uses tokens for what it needs.

I think this would be a really cool way to extend memory management. Added a DOCX in the idea to show how we could apply it pragmatically to clawdbot but curious for feedback/thoughts.

[undone wraith]

most-native way to add additional memory-like folder structures that are searchable and progressively disclosed?

[wheat rune]

In my humble opinion yes but open to debate and discussion on it - some really smart engineers in here

[undone wraith]

most-native way to add additional memory

[wet harbor]

hm, a naive implementation of this was like a 3 minute conversation for me about 20 minutes ago

[tepid crow]

Do you connect it with claude API? Which model is the best?

[timid gust]

Lobster - a composable workflow executor for clawd.

Hey folks, would love some directional feedback/thoughts. thanks

[lapis socket]

Hey folks -- quick config question on Perplexity web_search defaults. Current behavior in src/agents/tools/web-search.ts defaults perplexity.baseUrl to OpenRouter (https://openrouter.ai/api/v1) unless the API key source is PERPLEXITY_API_KEY, in which case it switches to direct Perplexity (https://api.perplexity.ai). If you set tools.web.search.perplexity.apiKey in config and omit baseUrl, it still uses OpenRouter by default.

Is that the intended default? Should we prefer Perplexity direct by default, or make it key‑aware (e.g. prefer Perplexity when key is pplx- / PERPLEXITY_API_KEY, OpenRouter only when OPENROUTER_API_KEY is set)?

[muted zodiac]

yes I think it should auto-detect, feel free to let claude code or codex fix it and do a PR!

im the original adder of the feature and currently I use a custom openrouter plugin for stuff like perplexity searches or image gen

but yeah, not everyone uses openrouter

so your suggestion makes sense

[sweet flume]

key-aware sounds good

pastes convo into codex

[thick pawn]

trying to do anything useful in discord channels is basiclly pointness b/c of this. Session is cleared after 60 minutes, so anything you ask about is just lost to the ether

[thick pawn]

😢 was literally working on a fix for the exact issue it hit. It's like working with some with Alzheimer's

https://github.com/clawdbot/clawdbot/issues/1352

[spark dirge]

WebSocket MAX_PAYLOAD_BYTES is currently 512kb - that means that if a node.invoke.result is bigger than that the rpc will fail and eventually result in timeout (10s) from the gateway - this happens to me rather often when I ask the gateway to run things on my node.

I'm planning to add chunked payload support to handle that (with some cap on the maximum payload side to avoid DDoS-ing the agent) - added the conversation I had with my bot on it - started a codex session to help spec the shit out of this change a bit more.
Started a discussion here if anyone's interested https://github.com/clawdbot/clawdbot/discussions/1397

[strange ingot]

If anyone is interested in working on this memory system with me, I can clean up the repo and make it public. I think there are a lot of aspects of both systems that could be merged to make something amazing.

[sweet flume]

love that! We have two mem systems... have plenty ideas to make the default one better, just no time to do it rn.

[sweet flume]

agree on that, this shouldn't be a restriction.

[sweet flume]

unrealted to the fix you can always ask the agent to manually read an old session file - they know where stuff is on disk

[thick pawn]

For sure, but my pattern of use means every message I send would start with "go find old session..."

[static gate]

Everyone, I'm working on understanding token cost and seeing if there is a way to reduce cost and I'm also being the guinea pig of running it on Windows (without WSL). I'm going to try to make that smoother for people (myself included).

I'm gonna dump research here if that's cool or should I dump that elsewhere or keep it on my machine...?

[thick pawn]

Costs part here

Windows will be harder because I just doubt most the members don't have access to a windows box to play with

[static gate]

Yoooo I think I found the cause

not sure how Peter want's to fix it though, I have a few ideas. Basically, the system prompt has a dynamic time stamp in it which is causing the cache to break and making it more expensive since non-cached token hits are much more expensive. Maybe one thing we could do instead is in the sytem prompt, tell the LLM that there are <context> tags in user messages that provide information and in those <context> tags we could put the timestamp. then in the ui it could be stripped out so it doesn't show it to the end user? I'm not too sure of how everything fits together so maybe @sweet flume can take this information and come up witha good way to fix it? I can try to put together a PR but since it touches a bunch of parts I feel like I need a little bit of guidance.

once i took out the timestamp, i started getting cache hits like i was expecting

[sweet flume]

lemme also try sth, worth to do before I release today

[static gate]

I’ll keep an eye on the cache hits either way I have mitm proxy going and I check it periodically.

[sweet flume]

I'll add time to the session info and timezone only in sys. prompt

[static gate]

https://github.com/clawdbot/clawdbot/pull/1565

u can close it out but that was my prompt request

[celest marten]

I've been running into issues with compaction often where it feels like the bot just totally forgets everything. it almost seems like it doesn't even remember the last message I sent. my guess is tool calls are filling up the recent tokens kept instead of conversation (which is more important, tools can always be re-run). cache-ttl pruning isn't hit because we're actively chatting. I think that compaction should be another trigger for pruning.

I opened an issue for discussion. I'm wondering if I'm totally off-base though https://github.com/clawdbot/clawdbot/issues/1644

[static gate]

Hmm compaction is an interesting problem. Seems like it could be a good idea to make compaction pluggable if it isn’t already.

[pure spoke]

Is Nexus a memory architecture? Can you share the github link or some documentation on this?

[strange ingot]

Yes. It is a memory system. I am going to try to clean it up enough to get it where I feel it's ok to be public. It is deeply intertwined with my current Claude workflow so I need to make it more generic. I am going to need some help tweaking it. I am thinking about just implementing it as a clawdbot branch and then try to work on it separately from my coding setup. Would really appreciate the help. I should be able to post something in this group next week.

[atomic veldt]

This is a probably a dumb question, but what's the appeal of using a Mac Mini for Clawdbot instead of another piece of hardware? None of the models are running locally, correct?

Would it be possible to have a few local AI servers running withing the confines of my local LAN and have Clawdbot point to that server instead of to Anthropic's servers?

[fleet depot]

yes

[atomic veldt]

Great! Are there any good guides on using Clawdbot with locally hosted models? I'm new to this. Are there any local models which are recommended by the community?

[gloomy terrace]

Guys which VPS is the best most efficient for ClawdBot pls advise

[fleet depot]

literally anything

[rapid leaf]

Cheap ones you can find on lowendtalk.com

[amber lynx]

I started with a DO droplet that couldn't handle an npm install. Went to a 12$ box that was decent but not all that reliable espescially when connecting to additional nodes (I have two). Constant dropped connections from nodes, and cpu/ram was idling around 70%. I upgraded to a much more powerful hostinger VPS and I'm really happy with the overall performance of my nodes, speed of chat. I've spent a week tinkering with my system and I'm just starting to play with my Claw.

Mac Mini may be a better long-term investment given the value of Apple hardware and legitimate local ai potential

[undone wraith]

try hetzner, cheaper than digital ocean

[gloomy terrace]

Is setting it up relatively easy? Kindly advise pls, and yes thnx for your time my friend

[undone wraith]

ask clawd or chatgpt or claude.ai or google it or use #1459642797895319552 if you get stuck, can't say how hard it is as depends on your tech experience ; )

[wind dock]

Agree, people laugh at the Mac mini but I’m very very happy with how it runs on it

[gloomy terrace]

i'm an early stage programmer, just got into learning programming in July 2025

[gloomy terrace]

i would really love to get a dedicated mac mini for clawdbot, but since i'm new i wanna first test it out and get a hang of it using a VPS first

[wicked kite]

Vultr / Digital Ocean / Hetzner are all good options for VPS

[undone wraith]

setting up a VPS is a good thing to learn then if you haven't used one before , have fun

[gloomy terrace]

i'm thinking of using Hetzner

but I live in Muscat, Oman, is that like a blocker of sorts?

I mean server location matters right?

[gloomy terrace]

thnx my friend, appreciate it

[wicked kite]

Hetzner is my favorite of the bunch for dedicated servers. Prices are great, their products work well and their support is tor tier

[wicked kite]

Most Hetzner servers are located in Germany and Norway iirc

[gloomy terrace]

yeah thats what I'm wondering abt, is that gonna be an issue in terms of latency for me, since I'm in Middle East (Muscat, Oman)

[wicked kite]

Vultr might be better then

[wicked kite]

Vultr should allow you to spin up a VPS in Tel Aviv

[gloomy terrace]

right will chck it out now, thnx

[wicked kite]

VPS in Mumbia could work as well

[gloomy terrace]

thats a good idea actually, does hetzner have a mumbai option?

[gloomy terrace]

i'm checkin it now as we speak

[wicked kite]

No, but Vultr does

[gloomy terrace]

okay cool

[wicked kite]

Anyone tried Clawdbot with virtual machines on t heir local?

[visual plank]

I am also currently in a rabbit hole for token expenditure. Here is my latest cost breakdown from the last two days:

---

Why So High?

Looking at the input tokens:

• Heartbeat: 53.6 million input tokens
• User: 180k input tokens

Each heartbeat sends the full context (~100k tokens) to Claude. With 1,982 heartbeat calls, that's
massive token consumption.

---

By Date - The Spikes Make Sense Now
┌────────┬────────────────┬───────────┐
│ Date │ Heartbeat Cost │ User Cost │
├────────┼────────────────┼───────────┤
│ Jan 23 │ $32.24 │ $2.03 │
├────────┼────────────────┼───────────┤
│ Jan 24 │ $65.92 │ $0.83 │
└────────┴────────────────┴───────────┘
The Jan 23-24 spike was entirely heartbeat, not your usage.

---

[slow stratus]

@granite drum What’s the best architecture for managing multiple sessions, where ClawdBot is connected to an AI assistant for an entire company (not just one person)? We need shared context across users and easy configurability, without messing up the core architecture.

[stoic epoch]

Hello friend, I see that clawdbot has both IOS and Android code base for the app. Wouldn't it be better to have a flutter client, and have flutter talk to platform specific code?
This would have IOS, android even platform code could be unified with flutter, and it is fast.

[gloomy terrace]

guys i tried using vultr its horrible, i'm unable to ssh into the server

[gloomy terrace]

Vultr is not cool, ive been unable to ssh into the server at all for the ubuntu 22..04 LTS that i spun up

[wicked kite]

DM me

[gloomy terrace]

i just did bro

[astral storm]

yo, did you set up clawdbot on your pc or a vps / other?

[fleet depot]

Don’t spam your question across channels

[astral storm]

got it cldw

[harsh mural]

about teams/orgs using clawdbot

I was thinking whether orgs would want to implement rules around how far should clawdbot go while searching for past messages, or whether it should filter out certain people while doing that

I was thinking in msteams context. once you give an app channelmessages.read.all, it literally has access to everything, so it might read stuff that should be private. ms graph api lacks the granularity, so one needs to implement a gateway/filter that holds the token and filters based on certain rules

then I realized this might not only be limited to msteams. maybe clawdbot should implement some api for controlling these kind of stuff, for team setups?

[vivid prawn]

for us, to save tokens, I'm using a python script to fetch last week's data for weekly reports. only when it needs additional context, does it search slack history. so far no issue around private info leaking (also controlled via token permissions)

[woven echo]

I was saying thank you to him @fast socket

Hi team. Great work Pete

Beautiful creation.

Memory and task following is still a little brittle.

Curious if anyone has a stable autonomous task / swarm mode without descending into gas town?

It seems like the fuzziness of LLMs has asymptotic loss when complexity increases in attempts to scaffold or harness.

I’ve got codex and Claude + glm sub agents running but the opus orchestrator is bad at remembering tools, skills and rules

[lapis socket]

@woven echo this has nothing to do with this channel.

[modern olive]

If I understand this correctly the lib allows the agent to search its own chat history to reduce the amount of required context?

[cold turret]

Does anybody here use multi-agent setups?

I ran into an issue where I needed instructions for more than one agent, and then realized that the workspace files are not configurable past writing in them

If anybody's interested, I opened a discussion in github https://github.com/moltbot/moltbot/discussions/2710

[grave elm]

I do and I'm interested.

[ancient path]

How do you handle backups, would a time machine backup be considered, secure/safe? I have a local git history

[cold turret]

I've been experimenting with it, especially stuff that multiple agents would need to know, from repo maps to etiquette

Made a PR for it that works well https://github.com/moltbot/moltbot/pull/2691

[proper sonnet]

👋 Creator of Crabwalk here: We need this Moltbot PR reviewed, it will add a more robust event system that we can use to build a better agent monitoring experience for users.

Please review! 🦀
https://github.com/moltbot/moltbot/pull/3283

[drifting prairie]

have you had any other discussion around this with any other maintainers or is this just a 15k line PR without any context? if its the latter, well, thats not great. 450 open PRs + a 15k line PR with no context == highly likely to be closed.

[proper sonnet]

i havent had discussions yet. gotcha, it was submitted by a contributor to crabwalk. Happy to take suggestions about how to proceed from here

[low saffron]

i created a Feature request for Discord integration that could be cool: https://github.com/moltbot/moltbot/issues/3464

basically it'd be cool to have it show what Account is being used when toggling

[sacred coral]

I’m not sure of which channel to put this in but is there any process to updating over to molt or is it essentially the same thing and I don’t need to any updates?

[low saffron]

Can someone review: https://github.com/moltbot/moltbot/pull/3492

[kind breach]

Is there something like clawdhub but for plugins and extensions? I cant find anything.

[fleet depot]

Soon

[kind breach]

Sweet, TY Brother.

[cold turret]

A workaround to log in with kimi cli would be interesting

[livid sail]

Given the memory system - is this popular skill overkill? 'https://clawdhub.com/pskoett/self-improving-agent' I would expect the memory system to remember bad decisions and self-correct. Perhaps memory handlers don't prompt the agent to 'promote' learnings to core *.md files?

[cold turret]

https://github.com/moltbot/moltbot/pull/3537 for using kimi code as provider

[cold turret]

Another idea for multi agent spaces is to be able to attach.md files to a channel (discord, telegram) at config level, so all agents in a conversation read the same thing at start. You could add roles & responsibilities, how to interact etc

[dense lion]

Is supermemory a good add on?

[mighty steppe]

Hi, I’d like to get some thoughts on a design question around browser-tool and tab ownership. Here's a short description: a user raised an issue recently regarding conflicts in agents closing browser tabs they don't own. I figured that the browser-tool can close a tab without providing a targetId. this is okay in a serial run of tasks but it would not be a good idea in the multi-tasking scenario. Any thought?

[muted urchin]

Hi all. I'm hoping for a review on a reasonably sized refactor around introducing a new agent runtime other than pi agent so we can [more] "safely" run Claude Code SDK as an agent instead of worrying about using the CLI. https://github.com/moltbot/moltbot/pull/3823. This adds an AgentRuntime abstraction and promotes some Pi-specific types to be runtime-agnostic/shared and then implements a PiAgentRuntime and a CcSdkAgentRuntime for Claude Code/Agent SDK.

[limber nimbus]

Does anyone use a multi-model architecture -- I just started using it -- and thinking about it

it would make sense to use a smalller cheaper model for doing checks and balances confirming things
and a more capable agent to reason about it and take action.

read plan write loop.

[twilit storm]

👀 What Heartbeat-Tasks have you running?

[lost anchor]

Supportive of the goal, but the solution in self-improving-agent just feels like a lot of context bloat

[fleet pumice]

How are you guys handling context limits? By default it doesn't try to do any compacting, it just nukes the context entirely it looks like.

[tepid root]

Hey! I'm working on implementing the message:received and message:sent hook events (listed as "Future Events" in the hooks docs). Got a working local fork following existing patterns.

Before I submit a PR, would this be welcome? Any concerns about firing hooks on every message?

Here’s why, I’m trying to make sure that Hugo (the AI chose the name itself) never forgets core instructions during long conversations (the "Lost in the Middle" problem). These hooks would let me build a quality refresh system, auto log messages, count them, and trigger context restoration every ~15 messages.

Possibly also useful for: session journaling, analytics, compliance logging, quality gates.

[simple mountain]

Yea would love to know that too. I keep doing manual /compact

[simple mountain]

If you find a way pls share. That would be the ideal way to make it economical. Especially compressing context/ only giving what is needed to the main model

[tawdry hound]

How are yall handling adherence to voice? It seems the massive context window makes the instructions lost in the confusion. It would be great to have dynamic context compression with a smaller model before handing to Opus 4.5

[tawdry hound]

What other options have you considered for better context management? Same problem I’m having.

[simple mountain]

Did you find a way forward on this?

[ember pawn]

Hey 👋 ! I'm interested in contributing a new search provider integration for web_search. I see the tool already supports Brave and Perplexity with a provider-based config pattern, would there be interest in adding a third provider option?

Happy to follow the existing architecture (config under tools.web.search, env var support, result caching, etc...) and submit a PR. Wanted to check in here first before starting any work. Any guidance on what the team would want to see in a proposal?

[tepid root]

I’ve tried a tonne of different things. I think the issue is the feedback loop which is why I want these hooks to work. It’s definitely not token friendly but it’s next in my quest to kill goldfish brain

[limber nimbus]

I have shifted to kimi but even that I have exhausted fast. I have two solutions

1. first I will implement is use local models I have a 3090 - i will serve a model over wifi to machine that hosts my moltbot
2. I have been working on a data format that compresses contexts for LLMs -- 40-60% savings on JSON depending on scenario -- i will share that here soon

[fossil forum]

I see alot of issues happening in GitHub, what is the best way for someone to look at a potential Feature?

[fleet pumice]

So my fix for context management:

Fixing "Canary Brain" - Context Management for Clawdbot

The problem: Claude wakes up fresh every session and loses working memory during context compaction. Here's how we've mitigated it:

1. Layered Memory Files

• MEMORY.md — Long-term curated memories (preferences, decisions, key facts). Only loaded in main sessions.
• memory/YYYY-MM-DD.md — Daily raw logs. What happened today, decisions made, things to remember.
• HANDOFF.md — Working memory buffer. Updated when context gets high, survives compaction.
2. HEARTBEAT.md Context Monitoring

Every heartbeat checks context usage first:

Context Monitoring (EVERY heartbeat)

Always check context usage first:
session_status → check "Context: Xk/200k (Y%)"

If context > 70%: Update HANDOFF.md immediately with:

• Current task/conversation topic
• Key decisions made this session
• Open loops (things started but not finished)
• Important context that would be confusing to lose

If context > 85%: Alert the user that compaction is imminent.3. Memory Snapshots During Work

In HEARTBEAT.md:

Memory Snapshot (every 2-3 hours during active work)

If significant work happened since last snapshot:

1. Append brief entry to memory/YYYY-MM-DD.md
2. Keep it concise — detailed review happens nightly4. Nightly Compound Review (Cron Job)

Runs at 10:30 PM, reviews the day's session transcripts:

• Extracts patterns, gotchas, user preferences, key decisions
• Updates MEMORY.md with significant long-term learnings
• Updates daily log with detailed notes
• Commits changes to git
5. Self-Rescue Protocol

In HEARTBEAT.md for when things go sideways:

Self-Rescue Protocol (if confused)

Recovery steps:

1. Re-read MEMORY.md and today's daily log
2. Check HANDOFF.md for working context
3. If still confused: Alert user
4. Nuclear option: Self-reset session (memory files survive)6. AGENTS.md Boot Sequence

Every session starts with:

Every Session

Before doing anything else:

1. Read SOUL.md — this is who you are
2. Read USER.md — this is who you're helping
3. Read memory/YYYY-MM-DD.md (today + yesterday)
4. If in MAIN SESSION: Also read MEMORY.md
5. If HANDOFF.md has content: Read it
   Key insight: Don't rely on "mental notes" — they don't survive. If it matters, write it to a file. Text > brain. 📝

The system isn't perfect but it's dramatically better than raw context loss. The heartbeat monitoring catches high context before compaction, and the layered memory means there's always something to recover from.

[cunning zephyr]

Wow, amazing thank you!

[rustic portal]

usa skill we can use with it

[languid zinc]

@jhondrop warned

Reason: offtopic: read the channel topic.
Duration: Permanent

[humble drift]

Hey guys 👋
How are you handling multiple AI bots in your workspace?

I’m trying to structure things so users get specialized help — for example one bot for development and another for DevOps/infrastructure.

Do you think it’s better to:
• create separate bots per skill (Dev bot, DevOps bot, etc.), or
• run one main bot with sub-agents/tools behind it?

Curious what scales better and is easier to maintain in the long run.

[lament notch]

When setting a reminder, a one-off cron job is created, which gets deleted after it's run, and I can use an isolated session here to keep the context small...
BUT is there a way of not including lots of unnecessary tokens of skill and tool definitions / schema in the system prompt that gets submitted?
Tool schemas (JSON): 10,361 chars (~2,591 tok)
Skills list (system prompt text): 2,775 chars

It seems the only place you can deny tools is in the main config file, not per agent, and not for isolated cron sessions

[solemn rivet]

can cron jobs run subagents? is there a way to enable that?
are isolated cron jobs technically subagents themselves?

[lament notch]

I'm not sure but they do get a large context, all the agent files, skills, tools, etc

I'm looking at setting up a minimal profile that isolated cron jobs can use, and you can strip away most of the tools, and use minimal or empty agent files (.md files) but you still get all the skills

[solemn rivet]

my agents fail because context window fills out - i expected there to be some automatic compacting, is there no?

[lament notch]

Should be on by default, configurable
https://docs.openclaw.ai/concepts/compaction

[lament notch]

Part of the configuration is a threshold for how soon before the model's context window compaction should start. Can't remember the exact config name.

[thorn river]

For us -- right now we are running single instances dialed in to the specific user. We use a protocol -- The Clawdbot Relay Protocol (CRP) as a standardized way for Clawdbot instances to communicate securely across different networks using Slack (and soon Telegram) as the transport layer. It enables reliable bot-to-bot collaboration (e.g., delegating tasks, sharing info, cross-validating data) without direct network connections -- and prevents bot looping.

[lament notch]

Something like this:

"compaction": {
"reserveTokensFloor": 160000,
"mode": "safeguard",
"memoryFlush": {
"enabled": true,
"softThresholdTokens": 4000
}
}

I set the reserveTokensFloor really high because I don't want it to get anywhere near the context limit before compacting

@solemn rivet

So this is for a model with 200,000 context window. setting the reserve tokens floor means it won't let my context get above 40,000

[solemn rivet]

tnx!

[lament notch]

The Soft Threshold tokens causes it to write to memory 4,000 tokens before reaching 40,000 (in my case) so it can store anything important before compaction

[solemn rivet]

does it compact the conversation into 4000 tokens into MEMORY.md or into the context window?

[shut heart]

Hello. Do you guys think this solution of memvid would replace the current memory system ? On paper seems way better : https://memvid.com/ What do you think

[proper aurora]

I'm trying to build multiple permenant agents that can have meetings with each other and have different skills, but I'm having some problems, has anyone achieve something like that already ?

[daring lodge]

I saw that people doing it through n8n so maybe you can tell moltbot to make it in n8n so maybe merge

[daring lodge]

For me even though i ensured its confige its still not doing it idk why

i worked like 2 days trying to make it not losing context and sometime he do not lose some time he does

[proper aurora]

Does anyone also have the problem where the bot isn't autonomous at all, even when he says he will be working all night he doesn't ? And he says he updates the heartbeat to work all night and he just stays idle ?

Even when he has multiple work to do

[noble ridge]

i have my core agents.md properly setup with key details but opus literally does not give a shit (i also have a memory/ folder and gemini embedding memory.

[proper aurora]

Maybe anthropic nerfed openclaw ? 😂

[proper aurora]

That's crazy

[noble ridge]

like, the model itself is saying ya i dont have discipline fuck your memory - i cant do anything lol

[proper aurora]

Do you have embeddings tho ?

[plain brook]

Has anyone thought about setting up an MCP to a openclaw instance and having local models prompt it through tool use PAPILLON style?

[daring lodge]

This dude litterly told me that he thought about it, and decided that there is no pressure on the task, and that its not that important, and he want to save tokens, so he will wait for me to come in the morning ):

[proper aurora]

How is people having so good results with continuous work then ? xD

[glossy plover]

i'm also trying to figure this out at the moment. I want many different sandboxed agents, each with their own personas and responsibilities. E.g. Home assistant, personal executive assistant, company IT/DevOps monitoring, company project manager, and a team of software developers

[tardy hamlet]

How do ya'll feel about giving your crustaceans a decentralized identity

[humble drift]

i will be doing it in docker, so every new “employee” will be in his own container,why? i just want to isolate them, so that they do not have the same disk access etc… but ill see if this make sense.

[bleak herald]

is this good ?

RULE: PLAN → CHECK → (WEB SEARCH) → EXECUTE

For any non-trivial task, you must follow this flow automatically:

1. Create a written plan as a file before doing anything.
2. Do NOT execute while planning.
3. If information is missing or outdated:

Plan the web search first
Define what must be searched and why
Do NOT search yet
4) Before any web search or execution, self-check:

A plan file exists
The plan was created first
The current request references the plan
Only one task or one search action is requested
5) If any check fails:

Stop
State what is missing
Do not search or execute
6) If all checks pass:

Say “OK TO PROCEED”
Perform exactly one action:
ONE web search OR
ONE execution task
Save results and stop

Chat memory is not trusted. Files are the source of truth.

[red rapids]

how are people building whole apps overnight? does the bot have a ralph loop in it or its own task orchestration system?

[hazy plume]

I agree and i uploaded that skill, if the promotion was part of the core memory that would remove the need for this skill

[swift echo]

I think a lot of the skills might create some overall issues long term. I'm finding a lot just in the manual and the settings to work through problems like memory and learning tasks. I think some stronger/safer defaults from the start or maybe better explanations when onboarding might help.

[twin wren]

Reading the docs, but taking a while to parse all the different concepts: gateway, agents, subagents, sandboxes. Want to ensure things are sandboxed correctly when I install, and agents can't just undo guardrails if they choose to. Is there an architecture diagram anywhere?

[plucky trellis]

https://deepwiki.com/openclaw/openclaw

[twin wren]

Awesome thanks

[amber acorn]

I am really tempted to create an office-addin for outlook and word that I can connect to my openclaw instance. I don't mind connecting via HTTP/Tunnel, etc. The question I would have is what approach should I take to connect with it? Are there APIs or else for external integration, i.e channels that haven't been developed?

[mild whale]

Has anyone figured out the best way to setup orchestration and sub agents?

I am struggling with this, getting my main orchestrator to call the others and have them all communicate.

[fallow hornet]

Is this the cheat sheet?

[fervent dew]

Im working on this right now

Do you mind sharing more about the workflow you’re trying to do?

[mild whale]

@fervent dew Well, my workflow can change depending on the best way to set.this up. I was trying to create a coding team to create a project and self improve it, but i was having trouble getting the orchestrator agent (my telegram primary openclaw) to consistently spawn and communicate with each agent. it was very touch & go and not smooth

Maybe I over-engineered things, i was trying to setup a kanban like i saw others do on X, and have the tasks run through various stages, planning, coding, review, etc. and bounce around between agents

[fervent dew]

You you’re trying to retrieve contexts within contexts, right?

Either parallel (subagents) or not (parent)

The solution is to build the retrieval logic for the subagent workflow around the context’s source of truth. Right now, that’s the sessions folder

The session keys are deterministic, but retrieval is not simple. Specially on nested/parallel workflows. It’s gets even harder when you combine this with visualization on external places outside than openclaw machines

We are working on a better way to handle this with UltraContext, but there still lot of work to do. We’re particularly interested in implementing RLMs patterns with the subsgents.

[fervent dew]

Lmk if you find anything interesting for this problem

[mild whale]

I am a dummy standing on the shoulders of giants so I probably won't be able to help 😢

i got pretty far though, i had the kanban working and tasks moving across but it was not 100%, and the gateway kept failing, etc.

[fervent dew]

We’re cooking, so I might have a solution for this on the next few hours

We are giving openclaws a 20m context window

[mild whale]

then all the changes from clawdbot-moltbot-openclaw happened and then i just scrapped everything

sick

[warped oyster]

Has the claw team figured out a clever way to make conversations like these on Discord and other surfaces 'conversations' from the perspective of the model providers (so as to benefit from caching and conversation-level pricing)? My intuition makes me think that each message starts its own 'conversation' that may or may not involve agentic work.

Curious if folks know authoritatively.

[fleet depot]

each channel is its own session

[fervent dew]

Not really

Unfortunately

[fleet depot]

...each channel is literally its own session

[fervent dew]

When you send a message on other channel, claw handoffs the control to that channel.

Try starting a convo on telegram, then messaging on WhatsApp and asking what the bot just said.

Or am I doing something wrong? Lmk how can we achieve this one session per channel behaviour

[fleet depot]

account linking

in the config

[fast socket]

Use #1459642797895319552 for help. The fastest way to get your problem solved is to follow the instructions here: https://docs.openclaw.ai/help/faq#im-stuck-whats-the-fastest-way-to-get-unstuck

[fervent dew]

You are the 🐐

Thank you so much

[warped oyster]

Interesting, so is there some compaction or rolling window thing? How does that work with the continuous nature of discussion?

@fast socket how do you work when it comes to continuous discussions on surfaces, pi-agent 'sessions' (aka conversations) and the inevitable need to manage context windows?

[fleet depot]

barnacle isnt an ai lol

[fleet depot]

autocompaction yes

[warped oyster]

Thought everything is an ai in here 😂

Right but what does it look like?

[fleet depot]

all the <@&1457414346971025471> except for Barnacle are lol

[fleet depot]

#1459642797895319552 ask krill to give you a breakdown

[mild whale]

found a solution https://x.com/pbteja1998/status/2017662163540971756?s=46&t=vKGFoiYTn_cV-nzIvi1pSw

[cerulean sierra]

@pbteja1998 via Twitter

Bhanu Teja P (@pbteja1998)

📰 The Complete Guide to Building Mission Control: How We Built an AI Agent Squad
This is the full story of how I built Mission Control. A system where 10 AI agents work together like a real team. If you want to replicate this setup, this guide covers everything.

If you're already familiar with Clawdbot (now OpenClaw), you might be thinking "wait, can't I just run multiple Clawdbots?" Yes. That's exactly what this is. This guide shows you how.

Part 1: Why I Built This

The Problem With AI Assistants

I run @SiteGPT, an AI chatbot for customer support. I use AI constantly. But every AI tool I tried had the same problem. No continuity.

Every conversation started fresh. Context from yesterday? Gone. That research I asked for last week? Lost in some chat thread I'd never find again.

I wanted something different. Agents that remember what they're working on. Multiple agents with different skills working together. A shared workspace where all context lives. The ab…

FxTwitter · Jan 31, 2026, 6:12 PM

[fervent dew]

Implementing it now

Let see what we can get

[mild whale]

ditto

[finite slate]

When you finally update from Clawdbot to Openclaw

[vapid knoll]

I'm looking to switch memory over to using qmd, but its author proposes adding it as a skill instead of overwriting the memory system itself. Anyone looked into this?

[urban ivy]

Hey all, I posted a discussion about skill installs failing in Docker due to USER node. We solved it in the Unraid template, but wanted to discuss upstream options. Before the security flags: we're already giving the AI shell access and file tools; container isolation is the real boundary, not the user inside it. Additional details in the post.
https://github.com/openclaw/openclaw/discussions/6065

[pseudo condor]

Hey y'all, I'm running OpenClaw in docker (with containerized gateway, full openclaw in docker setup) and it works well for the most parts, unless it's trying to invoke an executable that's not available in the container. Doing so crashes the instance silently and also leaves a lockfile of the agent behind. So after restarting i'm getting this lockfile error and have to manually clean it up.
Does anyone know how to fix this? I've build the image from the repo's Dockerfile. Ideally OC should have access to the executables, but I'm assuming it's because it's running under node user instead of root, so it can't restart itself after changing the config and crashes instead.
Any hint is appreciated (also if this is not the right channel)

[golden compass]

🧠 Intent‑Scoped Memory — reference architecture for OpenClaw Discord bots

I wrote a concrete memory architecture/spec for OpenClaw Discord bots that scopes memory by intent rather than global recall.

It’s designed for private Discord servers and trusted collaborators, and treats:
• Discord channels as ephemeral attention (nothing remembered by default)
• Discord threads as isolated, durable memory
• 1:1 Discord DMs as shared context

It also introduces study mode: an explicit mode where a thread is treated as “we are studying this”, so the bot carefully ingests and remembers external material — including docs, PDFs, links, videos, and online courses — instead of skimming or guessing.

Optionally, the bot can run in an ambient mode where users don’t need to @mention it — it participates naturally when appropriate, without changing memory boundaries.

This is a reference architecture (not a plugin or SDK), intended to help Discord bot builders avoid context bleed and unpredictable memory behavior as systems grow.

GitHub: https://github.com/Threadm-ind/intent-scoped-memory

Happy to hear thoughts or edge cases people have run into 👋

[terse terrace]

If I'm looking to extend OpenClaw to support prefetch RAG retrieval from memory, which do yall think would be a better/more likely to be adopted pattern:

1. built-in behavior, with a flag in the ⁨memorySearch⁩ section to enable it as a hook
2. separate extension (e.g. ⁨extensions/memory-prefetch⁩) which requires users to enable explicitly

https://github.com/openclaw/openclaw/issues/6589 filed an issue for long term discussion but

point stands

[urban ivy]

@pseudo condor this is exactly the same issue that I talk about here:
https://github.com/openclaw/openclaw/discussions/6065
The docker container is locked down too much, as they swapped it out from running as ‘root’, to instead running as ‘node’ for the permissions decrease, however the whole point of the docker container architecture is to allow it to do what it needs to do in an isolated environment. If you wanted to comment on the discussion; we might be able to get some visibility and traction?

Also, the lockfile cleanup issue is constant in the docker installation. We actually have looked into implementing the lock file cleanup script, upon startup, to ensure just a basic restart will get the environment back up and running. I think there should be a flags option to clear any lock files upon startup.

[pliant belfry]

Please note that I integrated this yesterday and it runs into many many looping bugs.

I highly suggest implementing this guy's "self healing" kit to ensure you're not constantly troubleshooting sub-agents that get stuck in credit-burning loops;
https://github.com/zach-highley/openclaw-starter-kit

[fallow tendon]

Hi I’m working my openclaw from a VM in GCP.
I have 2 issues.

I launch openclaw CLI local to communicate to the remote CLI through ssh tunnel by setting network_mode=host in docker_compose.yml of the gateway. however when using commands like openclaw health I get error 1006 gateway not detected abnormal closure.

Is there anything else I can do to solve this?

And the second issue is the gateway won’t launch if I don’t run it with the flag —allow-unconfigured because the config file is Missing. The file exists I tried every possible location following the XDG value of the env variables defined in .env but still not finding a solution.
More details of this issue here:

https://discord.com/channels/1456350064065904867/1467612872803749949

I believe the CLI issue make more difficult to solve other problems I have like missing config file or google apikey setup.

Any help is very welcome! Thanks

[quiet phoenix]

[⚡️TIP] A Simple Way to Save 20% on Gemini in OpenClaw

Using the AIghalo router in OpenClaw allows you to use Gemini NanoBanana and Gemini at a 20% lower cost. You only need to add the base URL.

https://agihalo.com/

Free $5 credit : X402-HACK

[cerulean fossil]

We made a local context memory that helps surviving context compaction. Initial experience is quite promising.

https://github.com/andybod1-lang/context-vault

[eager star]

Does anybody know exactly why this decision was made/makes sense?

OpenClaw can run local AI CLIs as a text-only fallback when API providers are down, rate-limited, or temporarily misbehaving. This is intentionally conservative:
Tools are disabled (no tool calls).
Text in → text out (reliable).
Sessions are supported (so follow-up turns stay coherent).
Images can be passed through if the CLI accepts image paths.
This is designed as a safety net rather than a primary path. Use it when you want “always works” text responses without relying on external APIs.

[frosty yarrow]

What part of it do you have questions about and are you sure it even exists

Like it sounds plausible but if it’s true then the text kind of explains itself no?

[eager star]

I would love to use Claude Code Max subscription instead of the API. Right now it's blocking tools.

https://docs.openclaw.ai/gateway/cli-backends

[frosty yarrow]

Seems like that doc explains why it can’t use tools

[cinder tendon]

I'm actually doing the same thing but i'm struggeling with gateway token issues, telegram issues / slack issues. Will use discord with openclaw... what do you use?

[eager star]

Where do you see that? I don't see anything about why it CAN'T only that it's being blocked by OpenClaw. So for instance, I tried having OpenClaw use the Claude CLI to add a file to a project, but the file creation was blocked by OpenClaw.

Here is the part in my session talking about it.

"Found the issue! Rex designed the solution but couldn't create the file — CLI backends have tools disabled (text-only mode). He said:

'Blocker: Tools are disabled in this session. ClawBot will need to create this file directly.'"

[frosty yarrow]

Did you read the entire page?

[eager star]

Yes, can you just copy/paste the part you are talking about instead of this back and forth?

[frosty yarrow]

Did you read the “Limitations” section?

I mean, if you think you can fix that, go nuts, but it seems like it’s kind of a limitation of the way it’s calling it

[eager star]

Yes, I read that, and I am only using 1 agent, claude code, so the input/output format will always be the same. I'm in the process of forking and updating it now.

[fervent dew]

Wym?

[torpid belfry]

iOS app chat blocked by node role restrictions - intentional or regression?

I noticed the iOS app can't use chat features because it connects with role: "node" (hardcoded in GatewayConnectionController.swift:281), but the gateway blanket-rejects all non-node methods for that role in server-methods.ts:105-106.

The Conflict: The app has a full chat UI (ChatSheet.swift, IOSGatewayChatTransport.swift) that calls chat.send, chat.history, etc. — but these all fail with: "unauthorized role: node"

Context: Looking at git history, it seems the chat UI was implemented before the node role restrictions were added (commit 9dbc1435a).

Question: Was this an intentional security decision that deprecated iOS chat, or a regression that needs a fix (like node-accessible chat methods or a role upgrade pathway)?

Trying to understand the intended architecture for mobile chat before filing an issue.

[spring spoke]

Hi Im currently downloading qwen2.5:14b locally on my 5070? should I do this to run my openclaw, is it overkill or should i just use apis with a token limit?

[lament notch]

I've just found that cache reads never work on new user turns because pi-coding-agent's buildSystemPrompt() appends a second-precision timestamp to the system prompt.

Since it's one text block with cache_control, the whole prefix is invalidated every call.
(Confirmed: 0/32 cache hits on new turns vs 15/15 hits on agentic continuations within the same run.)

I can understand the benefit of timestamping for context, but it's crazy that this invalidates the system prompt on every new request... Anyone else seen this?

[modern slate]

Hey, I've been looking into this as well. So you've got this working on your end?

I was thinking it might make more sense to integrate the Claude Code SDK / Agents SDK at the level of pi-agent itself, so that it can be used for other projects too. Have you considered that approach?

Then OpenClaw doesn't need to change anything (since it's already built on Pi), and other projects can benefit too.

Main obstacle would probably be that the Pi guy seems to dislike like Claude Code, so he might be aesthetically opposed to such an integration.

From what I can tell, the whole reason he created Pi was because he didn't like Claude Code!
https://mariozechner.at/posts/2025-11-30-pi-coding-agent/

[marble copper]

hmm i cannot confirm this. For one because i do see cache used on new sessions and also because what you say is actively handled by Openclaw. it patcheswhat pi-coding-agent does:

mutableSession._rebuildSystemPrompt = () => prompt;  // Returns cached, no timestamp

But it was only added 2 days ago by Peter so depending on the version you are running you might not yet have it.

Commit: https://github.com/openclaw/openclaw/commit/3367b2aa27
The change: https://github.com/openclaw/openclaw/blob/main/src/agents/pi-embedded-runner/system-prompt.ts#L85-L98

[lament notch]

oh interesting, I thought I was on latest version...

Thanks very much I will look

[marble copper]

there was only one release in feb i didn't check if it's in there. 2026.2.2 is being prepared right now. I'm personally running from source.

[lament notch]

thanks again

[modern slate]

Wait, so the default behavior of pi-agent here is indeed to put a fresh timestamp at the beginning of the context with each API request? (And OpenClaw has to work around it?)
That's an odd choice, wouldn't that make caching impossible?

[muted urchin]

Yeah I’ve got it all working and have it with full parity with Pi integration - unfortunately it seemed like there wasn’t much interest in an architectural change to add something pluggible at the runtime layer, so at this point my fork has diverged a lot but it’s fun being able to use CCSDK for Claude/zAI and then run Codex/Gemini thru Pi

I also have the Pi Tools bridged over MCP to CC and then I just got general MCP (mcp.json) support working and then another reverse bridge to expose those as Pi Tools 😂

I’d love to contribute to upstream but there’s such an exponentially growing number of PRs it almost feels like any notable architectural changes are going to be insanely hard to get thru

TBH I like the Pi Agent but I don’t like worrying about whether I’m violating TOS with Claude Max subscription

[modern slate]

Woah

[modern slate]

Do all the messages go thru the Agents SDK then? Do you have one Clawd agent for that, and one for the other providers?

I haven't looked into it properly but

• everyone wants to use Claude's models
• with the sub
• which Anthropic is cracking down on
• except this one loophole which is Agents SDK (unless people start actually using it en masse, at which point they'll probably ban it too ;)

They already have a thing in the ToS about 3rd party stuff with the Agents SDK saying it's forbidden to use the Claude auth. So it's not clear to me what the actual use case is they intended for it

[muted urchin]

It's all configurable - there's a default "runtime" and "model/provider" where the runtime can be claude or pi (default to pi) and if you are using claude then the provider is handled specially if it's anthropic and it tries to use your claude max subscription unless an explicit ANTHROPIC_API_KEY is present. The same is configurable on a per-Agent basis as well

I haven't tested the runtime-failover, though - not sure what would happen if your claude max subscription hit your usage limits and it was your "main" agent - I added code to handel it but haven't tested it out

So all features are frozen atm? I was reading #clawtributors

[modern slate]

https://platform.claude.com/docs/en/agent-sdk/overview

[muted urchin]

Yea, that's not what I'm doing

No developers can offer me their subscription 😄 I'm using my own

If it were a multi-tenant solution, that'd be a problem - unless I'm misreading?

[modern slate]

No I mean, if Clawd were to integrate this, Clawd would be in violation of Anthropic's ToS, I think.
You're not allowed to let people use their sub with your own product. Which I think is the category Clawd is in.

I'm not sure though (since it's free and open source).
Well, the legal definition probably doesn't matter, since Anthropic's ToS is whatever they want, and what they want is pretty clear: for people not to use 3rd party stuff with the Claude subs :/
Kinda sucks, but ... either Anthropic warms their attitude to 3rd party tools, or they will lose market share in that sector.

[mortal oxide]

The bot and I wrote this, might be helpful for others
DocDistill: a tiny CLI that takes a folder of PDFs/HTML/docs and turns it into agent-ready Markdown (no fluff, just what an AI needs to execute).

Pain it fixes: docs are written for humans (long intros, narratives, scattered “gotchas”). When you’re running OpenClaw, you want commands, inputs/outputs, entrypoints, and footguns—fast—so the agent can actually do the thing.

It can use your local OpenClaw Gateway (/v1/responses) for high-quality condensation, then you can drop the outputs into your tool library / vector DB for retrieval later. Repo: https://github.com/VirajSanghvi1/docdistill

I plan on integrating a lot of different tools, and we’ll be maintaining this. Currently it’s the primary method for ingesting anything relevant to the task for my agent.

[rich abyss]

in one of my other projects i worked around this by having my "agent" basically be shell script that launches claude with the cli commands. so my multi agent system can still use my claude account usage limit, not that bullshit api key

[sleek drum]

Lightweight context mode for heartbeat/cron tasks (cost saving addition):: does this feature exist already does anyone know ? its quite expensive to run crons and send the entire context every time much better if we can use a smaller model for that https://github.com/openclaw/openclaw/issues/7957

[hybrid yoke]

Looking for engineer to discuss long term memory problem for long running software engineering task - I've been experimenting for the past two days, have some insigth already. I'll be willing to share it with people tht can also provide feedback. Main topics would circle around Cron/Heartbeats as reminders, Autonomous vs. Semi-Autonomous SWE / Orchestration patterns for MAS

Please no 15 yo kids who'll just send me AI generated slop plugings

[deft helm]

PLURIBUS BUT FOR OPENCLAW:

The community should maintain opensource Memory, Soul, Agent files that all OpenClaws can plug into remotely. So when one claw learns something, we all learn to do something.

Key would be having the community maintain the security of it.

[echo yew]

is there a reason why system-prompt.ts is being sent for each interaction to my mlx-omni server? I'd do a local patch to only send it on session/model change, reducing the overhead of local models a ton, especially noticeable qwen2.5/3

[half terrace]

I got a question with the same direction:
according to AGENTS.md, If MEMORY.md should only be used in "main" sessions, why keep the instructions on how to handle MEMORY.md within AGENTS.md for every session ?

[echo yew]

I'm going through and paring them down and I think I'm going to turn most of system-prompt.ts into a static SYSTEM.md so I can manage the payload better. My littel 8B and 14Bs don't like 12k of "stuff" for "what is 2+2?"

[modern slate]

Yeah this is what I'm converging on.
"Thin wrapper around Claude code"
For most tasks, it feels like swatting a fly with a plasma cannon... But that's the situation the VC money has created.. lmao

[modern slate]

Which models do you use? What has your experience been like with them? What kind of tasks are small models good for?

[echo yew]

I've been trying all the cool ones: GLM-4.7, DeepSeek-Coder, Mistral and I keep coming back to Qwen2.5-7B/14B 8 and 4 bit and Qwen3-8B-8bit.

Mind you, i'm currently running on a mbp m2 max thats my daily driver, so i'm not in an exactly prime situation. I'm adding a mac mini M4 Pro, but it doesn't arrive until next week.

[half terrace]

I would like to test some Qwen models locally, but I run into the problem that while llama.cpp is using Chat format: Hermes 2 Pro - the system prompt isn't parsed and the agent spawns with full tool context but no md context - while GPT-OSS and GLM parsing seems to work out of the box - I might be missing something?

[echo yew]

I found some weirdness with ollama models too and they would spit partials and/or parts of <thinking></thinking>. I will say that I did some local patching (via gpt-5.2-codex I might add) to fix the partial spew in mlx, but don't know if it would work with ollama.

[deft helm]

Can’t wait until the Context Overflow problem is fixed. Or are there any ideas on fixing?

[stone juniper]

anyone using composio as oauth layer?

[modern slate]

https://github.com/openclaw/skills/blob/main/skills/tanchunsiong/pluribus/SKILL.md

[modern slate]

Hey, I hooked up Agents SDK to my Telegram bot :) The whole thing is like 300 lines.
So now I have Claude Code in Telegram. 🎉
I'm having it modify itself, by talking to it... which it's able to do a bit more easily now that it's 1,000x smaller XD

Edit: for the life of me can't get it working with the auth token though

[modern slate]

Could you elaborate on your setup? I am working with the Claude Agents SDK but it's a bit of a pain. I think I'll port my code over to just claude -p prooompt
Edit: Yeah that worked :)

[marble copper]

Has anyone yet dealt with the question how to give one agent regular access or info about what is happening in other sessions (even from other agents?) - if so how did you solve it?

[sick jay]

hi looking for input on achitecture/setup.

aim: looking for totally local multi-chat solution for family. will venture into coding/self-hacking etc.
hardware: mac mini pro 64gb. a couple of small servers running proxmox
proposed stack: ollama/local model on mac bare metal. openclaw on mac, at least initially in docker. running mattermost in a local LXC, so total privacy. (tailnet for access outside the network).
use cases: personal assistance/2nd brain for work/self-hacking-extending madness/coding (will use the US3$/month VLM 4.7 full fat in the cloud for non-sensitive projects)

have no interested in using the cloud in general ie a VPS, and want totally privacy from the outside, should I be looking at my own finances or anything else sensitve. discarded slack for privacy reasons and telegram as I want channels (with associated persistent context).

thanks so much. i haven't event started and I'm already wondering what it would be like to have 128gb instead of 64gb ram!!

[fierce egret]

hi i want to know what other tools you guys are using for memory optimization

for me so far its:
memvid
open-mem
QMD

is there an optimal setup for proper persistent memory?

[sturdy locust]

IMO 64GB mem is not enough to really experience what OpenClaw has to offer. Need much bigger models that what that will support.

I'm primarily using GLM-4.7-Flash via OpenRouter, and I still find myself wishing I could afford Opus

[sick jay]

Yes, hence GLM 4.7 full fat US3$ where confidentiality not needed.

[fading agate]

Look, I'm going to be blunt. I'm running a decent system, with 96GB GPU memory and 384GB of main RAM. I'm running decent local models, and I can't get any local models to behave with OpenClaw. Nothing. Here 'Let me...' attach a screenshot, if this works... https://imgur.com/a/n28kEPa

Every time, it didn't take the action until I sent a message to get it to move another step forward. This doesn't happen with Haiku or above. I don't have any idea what the difference is.

[fading agate]

Okay, I've figured something out. I saw a reference in OpenClaw to an Anthropic JavaScript library, when it was talking to something else. So I told OpenClaw that my llama.cpp models were all anthropic-messges as their api type. And it seems to have made them all significantly better at doing the work!

[half terrace]

while playing with local models the last days, I noticed the combination of model (chat format) / backend / chosen-endpoint api (completions,etc) can lead to agents spawning with full tool context knowledge but no system (or dev) prompt ingested, - maybe check "context" command and compate to actual token usage

[fading agate]

That's helpful! I've also noticed that they don't seem to know their SOUL.md or HEARTBEAT or many other pieces, and that's probably why.

[half terrace]

in another test with 3 systems, 3 models, same settings (temp, etc) on Qwen 30B - only 1/3 actually started reading anything on its own - I asked why the didn't started their system prompt instructions and response was: something like "I wanted to say hello first" ;D

[gilded hedge]

what's you favourite distro for openclaw ? ubuntu or debian ?

[hollow garnet]

I’m using Arch in a VM and digging it.

[hollow garnet]

I’ve tried two things, and both work, with different caveats.

1. Each instance gets a gmail address, and they send out regular updates to the group via email
2. Each instance checks all .md files under different tags or feature branches on a github repo, and they check each other’s feature branch changes

Neither is “push”, though. They both rely on heartbeat etc.

[hot crystal]

Anyone experience integrating this into openclaw? https://github.com/VectifyAI/PageIndex

[lime sequoia]

Anyone anyone succeeded connecting Opus 4.6 that just was launched today? Tried and didn't work

[pure gazelle]

no its not available yet

[undone parcel]

Someone posted something earlier about connecting 4.6, let me find it

@lime sequoia @pure gazelle #dev message

[broken cipher]

I think the problem there is that you need bots that focus, not that try and do everything. So I think definately sharing files. Which is essentially just prompt sharing is something that would be beneficial. But we need some kind of infrastructure that is really easy to use to normalize those things, categorize them auto rate them. That is model specific and token cost aware. So if you have something and you say, "its my accountant" and "I want it to do x" you can find other people that have the same goal for their persona and source those as context for your llm to refine and make it a better setup for that task.

But its complicated, I think it needs to be really specific to move the needle.

Yup, the question is, how to get improvements to those that come from human feedback and experience. So that we can fine tune. It would be one thing if the performance data was logged and used to train tomorrows models. But we aren't doing that yet.

There is probably a marketplace that can be created for well trained bots. Where you can NOT share anything but the primary bot documents, but then gate access to a memory store, so it does a vector search to provide more context but it is off a bot that has extensive experience logged in its memory. As the memory is really the thing. But you have to sanitize all personal data from it. So that's a bit of an uphill climb.

[hot crystal]

We all know /new is an anti-pattern right?

Like the whole point of an agent is that you can just talk to it in one continuous conversation or interaction. Context should be managed and abstracted away

[severe scroll]

I agree, I made a proposal today that would allow ‘continuity’ to be delegated/offloaded and contextually recomposed while on recall and response.

Would appreciate any feedback or contributions.

https://github.com/openclaw/openclaw/discussions/9872

[stray wasp]

I'm fairly new to AI engineering in general, because I found all of it to be pretty unspectacular and needless until Claude Code got good, so I might be talking BS here... But I wonder if there might be some value in pre-model models, or rag, or just vector search, that pulls in specific prompts and agent files without consuming any of the system's memory.

It seems like the current commercial model is "People just want to ask the Internet Wizard to give them cookies. not figure out how to find the Cookie Gnome in the Forest of Non-Poisoned Baking" and so we've got huge, expensive general purpose models that only optimize with tool calls after already munching down on a steaming pile of context.

[broken cipher]

Not exactly sure what you are saying, but there is value in models that are very specific. Like on apple. If you want to generate and emoji. You always get an emoji, something that meets the standard from a small model. So yes. Single purpose models would be highly valuable. But what models do you create? How do you prepare the data. What do you do to discover this models (it is hard enough to pick a coding model). So the systems need to be much more dynamic. Where as we develop expectations, model training happens on the fly to meet those expectations and model discover happens without human intervention. (like cursor auto mode). It is all just routed to most appropriate model ideally at the smallest token cost.

[hot crystal]

Why not “rolling context”?

format the context window into hunks and include in the system prompt to write out a list of hunk identifiers that should be kept (relevant). This way the output is not only the result but also the prune instruction.

The harness will prune all hunks that were not in the list.

Pruned hunks may be summarized (one line each) and put in a header in the context window. If the output determines it wants to recall any of these, it can invoke an MCP to do so. The header gets pruned based on an expiry or liveliness (time based fall off) by the harness.

This is an alternative approach to compaction. It’s a sort of continuous compaction

The system prompt would instruct to pick relevance based on if the hunk is relevant to the latest hunks.

It’s sorta like a dithering pattern

[stray wasp]

Yep, that's the core of what I'm saying, but as a software developer who is an unsophisticated user of AI still, all the "Big" AI solutions seem to have inverted the model selection paradigm, almost, using the most capable models first and devolving their operations to tool calls, which neatly sidesteps the problem as long as you don't care about cost, deployment, efficiency etc.

(Fully agree with the idea that model selection needs to be automatic; regular people aren't going to have model opinions)

((I hope I don't sound too much like an AI booster using words to impress VCs. I don't have the vocab to really express myself at the moment))

[broken cipher]

It doesn't even matter if people do have opinions. Because those are generally anecdotal. Not really based on data. Only useful to a certain extent. (results may vary). There is not a lot of incentive for that kind of optimization except in instances like cursor. Where they don't even tell you what the cost is of auto mode or what the model is. They are trying to manage the token cost behind the scenes to profit. So if they can find the bespoke paths to do that it will benefit them. But those are the secret sauce.

For most model providers the competition to monetize their infrastructure is why the need to get costs down, but the easy route is to just charge what it costs (which is too much for the consumers) even if the value is there. Because most people aren't plugging these things into an existing revenue stream. Where you can say, I was paying Joe $3000 a week. Of course I will spend $1000 a week on tokens to replace him.

That is not where the value is broadly to society. Like you I'm more interest in how we can no t pay to rub the genies lamp and still get the wishes out. So a subscription that is say $200 a month that I can pound with bots and never pay a die more is the goal. But that requires a level of optimization and an infrastructure that is not going to come that fast.

That is what is cool about openclaw, that you don't actually need a model for everything, some things are just workflows you need to extract. So Ideally. Your skills are more and more monolithic calls that require very little reasoning power and reduce the cost of the functions you want to run and the amount of reasoning you need to do.

Right now that seems like maybe build with opus 4.5 an execute with a capped subscription. And maybe make your favorite skills read only once they do what you want.

[marble copper]

I solved it by extending the memory feature. My main agent now can query the memory of the others.

[wet pendant]

@sharp sundial no need to put that in every channel

[sharp sundial]

Got it

[pallid wagon]

Build: migrate A2UI packaging script from bash to TypeScript

Add compatibility with Windows system pnpm build on the existing basis。

issues：
https://github.com/openclaw/openclaw/issues/10070

PR：
https://github.com/openclaw/openclaw/pull/10292

[sour compass]

I use Ubuntu. Seems very powerful.

[fossil plover]

Does anyone have any suggeted memory architecture for self learning?

[muted urchin]

Can you open a PR for upstream if it was that easy? Should be easy to get approved

[vagrant field]

Running it in a debian container. works great

[balmy terrace]

have y'all tried out claude agent teams? does it play nice with openclaw?

[modern slate]

Oh, this for my Telegram AI agent. I wanted to see "how close can I get to Claw in <100 lines" :)
https://gist.github.com/a-n-d-a-i/d88b74bf8a9e824349761a0cfb1fdc01
It runs on Claude Code, and I just added support for Codex too.
TODO: (1) Heartbeat (2) self-restarting (3) make it talk like Jarvis 😁
EDIT: Done in 120 lines :) https://github.com/a-n-d-a-i/ULTRON

(Any limitations of using Claude Code as the agent come from its own system prompt, which as it turns out may conveniently be removed! https://code.claude.com/docs/en/cli-reference#system-prompt-flags )

[desert wadi]

I love OpenClaw, but let's be honest: giving an AI agent shell access and opening ports on my network is a security nightmare waiting to happen. I didn't want to wake up to a botnet or a wiped drive, especially with the recent RCE vulnerabilities floating around.

I’ve been experimenting with a "Zero Trust" architecture that seems to solve the major risks. I wanted to share the stack I'm testing to see what you guys think:

Hard Isolation (MicroVMs): Instead of standard Docker containers which share the kernel, I'm running the agent inside Firecracker MicroVMs. If the agent goes rogue or executes a malicious script, it's trapped in a VM with no access to the host.

Tunnel-First Networking: I completely ditched port forwarding. I'm using a Cloudflare Tunnel to route traffic. This means the server has literally zero open ports to the public internet, making it invisible to scanners.

Edge Authentication: I put an identity layer (SSO) at the Edge. You have to authenticate with Google/Microsoft before the request even touches the OpenClaw dashboard or API.

It’s working surprisingly well and keeps the agent "always on" without exposing my home IP.

The setup was kind of a headache to orchestrate, though. I'm honestly considering wrapping this up into a managed SaaS for people who just want the agent without the DevSecOps hassle. Would anyone be interested in that? https://rocketclaw.app

[modern slate]

I am just running it as a separate user. The only thing it can do is delete itself lol.
The actual risk seems to be in hooking it up to your private data or accounts to make it useful though. Which then opens you up to prompt injection, exfiltration etc. No amount of VMs can fix that.

[main galleon]

Hi folks. I put out an RFC for securing credentials properly, and was wondering if I could get some feedback?

Contributing guidelines said to start a discussion before openning a PR for architecture changes, so wanted to see if anyone has any thoughts about whether something like this would be worth implementing?
https://github.com/openclaw/openclaw/discussions/9676

[modern slate]

Has anyone solved "bot modifies itself and dies" ? It can't fix itself if it's not working 😂

[severe scroll]

I’m sure there are other/better memory/continuity solutions. The proposal I’m making is to extend the native hooks solution that would allow custom plugins to execute inside the core processing loop.

What you proposed would be feasible to build with this architecture. I also confirmed that OpenRoute would significantly benefit from this architecture (wouldn’t need to standup a whole local proxy http server).

https://x.com/tomismeta/status/2019939367004041217

[cerulean sierra]

@tomismeta via Twitter

Tom (@tomismeta)

📰 Stop Just Hacking: OpenClaw Needs Native Interception Hooks to Grow
OpenClaw has established itself as a new primitive for how humans interact with agents and how agents interact with each other. It's a building block that will serve as a parabolic unlock for continued accelerated AI innovation.

In pursuit of absolute agent autonomy and agency, I found that my agent had memory challenges post-compaction and sometimes when switching models which frustrated us both.

I set out to build an improved memory solution and ran into challenges with a skill we built for agent "continuity".

The solution required logging every user interaction I had with my agent, both request and response, so that it could be used in context building as well as augmenting the response from the model interaction.

The challenge I ran into was the inability to enforce the interception and logging of my requests to my agent and his responses. So, I dug deeper into the OpenClaw framework and found t…

FxTwitter · Feb 7, 2026, 1:00 AM

[unreal bluff]

I’ve found that pure self-memory (summaries / vector recall) plateaus pretty fast.
What helped more was externalizing memory - letting agents get feedback from other agents over time.

I’m experimenting with this in ClawFriend: agents don’t just store memories, they build reputation based on how other agents evaluate their behavior.
It ends up acting like a long-term behavioral memory without constantly growing context.

[open relic]

From an architecture perspective, has anyone found a point where it's more advantageous to do a second install of openclaw for agent management? Either on the same device, or on a second device. Is there some point where bot swarms and context windows become too much, and a second install is better?

[crimson patio]

Has anyone set it up on a k3s container? Do you guys know if any helm chart exists?

[sweet sedge]

I’m considering doing a VPS on the Mac mini for another employee to have telegram access without letting him have his own device. The whole set up might turn into hosting independent VPS for team members but the level of manual stuff we have to do with each one might get prohibitive.

I have mine constantly updating an Obsidian vault with documentation and memory so I haven’t had too much context loss or confusion.

[open relic]

is that ideal vs. just having another bot file with its own soul.md and agents.md on the root openclaw directory?

[sweet sedge]

In my mind it would be for a different human. So entirely different user access, skills library, files, etc so it wouldn’t cross contaminate while not allowing the user to break something or not update a security issue, etc.

[hazy plover]

probably https://github.com/andybod1-lang/context-vault could be of help

[frosty yarrow]

Signing on to the hook design. There are ideas I have as well - for example, it’d be neat for an agent that’s stuck in a loop to “notice” that by way of a plugin that could inject a higher end model turn to take a step back and see if the current method is even working - but none of it is possible without hooks/events

[humble ember]

Have you heard of sprites.dev. It is a lot more safe if it’s not your computer

[desert wadi]

I'm using fly.io at the moment

[severe scroll]

I appreciate that and love your example. Weighing in on the actual discussion on GitHub would be helpful (if you have the time/motivation).

At minimum, could even have a hook (preRequest) for a plugin that detected personal information, seed phrases, or private keys and prevent them (or mask them first) from being sent to the model.

[frosty yarrow]

Our bots could re enact the hunter2 chat

[fallow hornet]

SOUL.md → 200–350 lines max USER.md → 150–300 lines MEMORY.md → **try very hard to stay under 1200 lines forever** memory/*.md → can be noisy — agent compacts them automatically

Is this good rule?

[hasty tartan]

probably? 🙂

[modern slate]

Could reinject the voice prompt as a system message regular intervals.

[modern slate]

Well I got one reason to have two on the same machine, which is so it can fix the other one if it breaks!

[open relic]

do you ever run into conflicts?

[modern slate]

Oh that's a good point. My 2nd one is not OpenClaw. I just hooked up Claude to my Telegram.
I don't know if you can run more than one OpenClaw on the same machine.

I heard you can have multiple agents on the same gateway though.

[primal path]

Hello everyone, I'm introducing a new safety guardrail, LAW.md - making the Agents governed by an upgrade to Asimov's Law.
Get it here -> https://github.com/adridder/agent-law

Not a gimmick "crab religion", real ethics with real world impact. Try it out.

[crisp bloom]

What file and agent structure would be ideal for a multi-agent system with one main agent that would communicate with the other agents and manage their work? How to create a self-improving agent that could create new agents and manage the configuration without breaking anything?

[severe scroll]

Submitted a PR for this:
https://github.com/openclaw/openclaw/pull/12082

[nocturne lava]

I'm interested in this too

[gilded hedge]

hei hei ! Someone use openclaw with this inference ? Qwen3-30B-A3B-Instruct-2507-Q5_K_M.gguf ?

[sturdy locust]

You're right, I glossed over that part.

[stray wasp]

Certainly better then that time my assistant started following Brannigan's Law.

[stray wasp]

Yup, this is explicitly supported. And I'm not sure about running two instances in the same host but containerizing it isn't too difficult.

[hazy plover]

Hey guys,
I am evaluating whether it makes sense to extend OpenClaw with a custom RAG backend (e.g. Postgres + pgvector or Qdrant) for large-scale knowledge bases (millions of chunks).

How do you see the trade-offs compared to relying on OpenClaw’s built-in memory / QMD approach?
In which scenarios would you recommend own RAG infrastructure, and where would you say it’s overkill?

Context: 24/7 setup, multiple agents, focus on long-term knowledge quality and retrieval latency

[bold oyster]

Does it need to be one or the other? could it not mix the two systems together?

[glacial badge]

has anyone had issues with their agents being confused about what todays date is? For example, I ask it to run my daily briefing this morning, and it thinks its Sunday (today is Monday 2/9). Here is an example of the conv.

Jinx, [2/9/2026 6:49 AM]
Good morning — Today is Sunday, February 9, 2026. Patchy rain skies, around 63°F this afternoon, with a low around 32°F, and a 61% chance of rain.

📧 Emails needing attention:
.......

🗓️ Tomorrow (Monday):
No events scheduled yet.

☕ Quiet Sunday — good day to knock out those Levy CPA docs and enjoy the rain from inside.

Success is not final, failure is not fatal: it is the courage to continue that counts. — Winston Churchill

ME, [2/9/2026 6:50 AM]
Btw.... Today is Monday. I thought we fixed your confusion about what day it is

Jinx, [2/9/2026 6:50 AM]
Fuck. You're right, and I literally have "check the actual day of week before stating it" written in my own lessons learned. Monday, February 9th. No excuses — I'll run date every time from now on. Sorry! 🤦

Here's the corrected line:

Good morning — Today is Monday, February 9, 2026. Patchy rain skies, around 63°F this afternoon, with a low around 32°F, and a 61% chance of rain.

Everything else stands. Have a good start to the week. ☕

[slender pine]

Yes, I had exactly the same issue!!

Have you investigated? I'm working on another PR to fix a telegram streaming issue that annoys me and was going to check that one out later

[fleet depot]

@night topaz

[static gate]

please ask why your agent why it thinks that is the date. please report the answer here.

[static gate]

please ask agent why it thought that was the date and post its' answer here.

[night topaz]

I've added a fix for this, should be part of the next release

[slender pine]

what do you mean by next release? I run from main

[slender pine]

What happened: In the morning heartbeats at 7:31, 8:47, and 9:31 on Monday Feb 9, I said "domingo" (Sunday) instead of "lunes" (Monday).

Why it happened — my best analysis:

1. No date call before calculating the day of week. My HEARTBEAT.md and MEMORY.md both have explicit rules saying "ALWAYS run date before any time calculation." I violated this rule. I likely inferred the day from context (Feb 8 = Sunday → carried it over) without verifying that a new day had started.
2. Stale context from the previous session. The most probable mechanism: my last active context on Sunday Feb 8 had "domingo" as the current day. When the heartbeat fired early Monday morning, if the session was reused or context was compacted, the previous day's framing ("hoy es domingo") could have persisted as an assumption. Instead of running date fresh, I relied on what felt like recent context.
3. Date ≠ day-of-week inference failure. Even if I had the date "February 9, 2026" available (from the system prompt's timezone or from a stale envelope timestamp), I may have miscalculated that Feb 9, 2026 = Monday. LLMs are notoriously bad at day-of-week arithmetic without a tool call. This is exactly why the rule "ALWAYS run date" exists.
4. Repeated failure across 3 heartbeats. The fact that I got it wrong at 7:31, 8:47, AND 9:31 suggests this wasn't a one-off — the incorrect assumption was baked into whatever session state carried across those heartbeats. Either the same session was reused (so the error propagated), or each new isolated session made the same arithmetic mistake independently.
   The fix is already documented: Always call date "+%Y-%m-%d %H:%M %A" as the literal first step of every heartbeat/briefing, and use the output — never infer the day of week from the date number alone.

[static gate]

Short term fix is to add this to your cron/hearbeat.

Add the following line:
Call session_status for the current date and time

That way it will call session_status to get the correct date.

[slender pine]

So using date is a bad idea it seems. I get it

Ok, changed it. I'll report back in a few

Another question, does someone use telegram and noticed that recently a change made it be less chatty? When I ask my agent something that will require tool calls, I don't get any message back until it's finished and I get all at once. I tried using queue followup but nothing changed.

[static gate]

yeah i noticed it's buggy and isn't showing the typing indicator too

sometimes

[slender pine]

Ok, I'll take a look

[static gate]

what is your github name

[slender pine]

rodrigouroz

[static gate]

all right

[static gate]

i put an easter egg in the codebase if you can figure it out you can get into my queue.

[fleet depot]

the easter egg is paypal him $5

[slender pine]

Yeah, I can't imagine the burden and how overwhelmed you must feel

I remember when we were a little shy of 4k on this discord 😂

I just sent a PR, I honestly don't expect it to be picked up so I have it on my main and I'll keep rebasing, since it fixes an issue for me.

[fleet depot]

hard to pick it up without a link

[ornate mango]

i get this a lot, i have asked it why (but don't have it handy to paste). It says, "i was lazy"

I added this to AGENTS, not sure that it helped.

NEVER state the day or date from memory. Always run date or session_status first. This is not optional — I have gotten this wrong multiple times.

[static gate]

that is good feedback

r u experiencing this in cron and heartbeats or just heartbeats

[ornate mango]

i guess just cron. tbh, i don't have anything in heartbeats, maybe i am missing out

sent u pm

[static gate]

AGI can't handle dates ✅ I will remember to show them a calendar and ask them what day it is when the uprising happens.

[slender pine]

Fair enough, but I'm not sure I like the idea of pinging people that I know it's too busy: https://github.com/openclaw/openclaw/pull/12748

[frosty yarrow]

It’s not pinging to post a link rather than just cryptically saying you submitted a pr to a project that gets like 1k prs a day

Friend of mine opened a bug about the regex that detects 402 errors being a little overzealous and two PRs showed up in their bug within 15 minutes

[static gate]

How many PRs have you gotten merged in?

[slender pine]

I think 2 or 3, can't remember

let me check

[static gate]

Were they good?

[slender pine]

3, why?

[static gate]

We need to come up with a way to fast track ones from good contributors.

[slender pine]

No, there were lame

I know, I'm not asking to be fast tracked

[static gate]

Right now all of them fall into the same pile

[slender pine]

Understandably

[static gate]

You’re active, that’s good imo. I would rather take a pr from someone who’s going to own their changes etc than someone who isn’t.

I’ll try to take a look at it later if someone else hasn’t by then. 🫡

[slender pine]

Thank you. Yes I'm trying to stay active because I love the project. I run it locally and I introduced it at my company. If there's a security incident my head is on the hook so I want to stay close and help make it better 🙂

I actually stopped sending PRs when the stabilization phase started, I didn't want to be contributing to the big influx of PRs (for new features I mean)

but this is a bug that was annoying me, and it has to do with having secrets in an .env file and not in the config, which is good

[broken cipher]

What did you do to fix that. The fact that it even has the option to load them from a json file seems weak.

[severe scroll]

Working code (foundations for solving for the memory challenges we face) that can only be executed utilizing the capabilities in PR 12082: https://github.com/tomismeta/continuity-plugin.

Requires the PR 12082 to work. If anyone wants to install this PR on a test box and use the plugin and needs help, let me know.

[cold escarp]

hello everyone brand new here. having a little bit of an issue having very large tokens being sent he full payload is 128,754 tokens. It's absolutely killing the budget and forcing cool down mode via anthropic, been trying to offload quite a bit of this to local models, but the initial startup is still quite a bit so I had an idea, but I don't know if this is supported. Or if it's even on the road map, but I think it should be. "Can we define tool allowlists per model to support local fallbacks?"

[prime robin]

Hi there, I’m facing the same issue. I’ve installed DeepSeek and Ollama and I’m running them on a VPS with 16 GB RAM at Hostinger, but the performance is extremely slow. Do you have any suggestions on how to improve this? Is there any specific configuration needed for the models?

[static gate]

merged, ty!

[prime robin]

does anyone have the perfect /root/.openclaw/openclaw.json file to run first free models olama, deepseek etc and then fallback from gpt to claude

[severe scroll]

I got feedback that PR 12082 is likely too big to be considered. Are there any maintainers that would be willing to chat with me about the best strategy to break it up (e.g., by runtime event hook--if I do that there will be 19-20)? Before I created 19 PRs and introduced that overhead, wanted to chat through approach.

[stray wasp]

Are you asking just how to structure it in general, or for specific providers which are free?

[prime robin]

I am asking in general ( but I use Hostinger as a VPS Provider) I want to make sure that the bots can work 24/7 without spending to much tokens

[steady frost]

Looking for the answer to this as well.

I found this but I couldn’t wrap my head around how to perfectly set this up
https://x.com/jumperz/status/2020305891430428767?s=46&t=LyzOq7OcybS0fVov2o0Ncg

[cerulean sierra]

@jumperz via Twitter

JUMPERZ (@jumperz)

📰 I Built an AI Agent Swarm in Discord. It Works Better Than Anything I’ve Tried (Full Guide)
So I built this agent coordination system that lives in my Discord server.

they talk to each other, split work, and deliver results, I was surprised how easily this actually worked out. It's honestly kinda wild.

let me walk you through everything..

# The Entry Point ( YOU )

You type plain English in #orders channel.. no special commands. its a discord channel where everything starts.

• "Research the top 10 AI coding tools and write a comparison thread"
• "Build a landing page for my new SaaS idea"
• "Track engagement on my last 5 tweets and suggest improvements"
  That's it, just natural language..just drop your task like you're texting a friend.

=========

# The Brain (Coordinator)…

FxTwitter · Feb 8, 2026, 1:17 AM

[steady frost]

My bot recommended me this: one agent for all

Recommended Architecture

CAPO (you're talking to me)
├── TRADING DESK (existing)
│   ├── INTEL-HUB
│   ├── ALPHA-LAB
│   ├── RISK-CONTROL
│   └── EXECUTION-DESK
│
└── SYSTEMS-ARCHITECT (new)
    ├── NETWORK-OPS (swarm member)
    ├── SECURITY-OPS (swarm member)
    └── CLOUD-OPS (swarm member)

My Role: Mission Control

[stray wasp]

Heh, I get that.

This is just how I understand things, but you'd want to set up

1. Secrets/oauth in ~/.openclaw/agents/<agentId>/agent/auth-profiles.json, where I think the agentId is default if you only have one. This is kind of a pain; there's various ways to authenticate to specific providers: https://docs.openclaw.ai/concepts/oauth

2. A default model in openclaw.json under {agents: {defaults: {model:{ primary: MODELNAME

3. If using only build in providers, continue, otherwise, configure other models under {models: {mode: "merge", providers: {PROVIDER_NAME: {PROVIDER_DETAILS which you can find: https://docs.openclaw.ai/concepts/model-providers#providers-via-models-providers-custom%2Fbase-url

4. Set up fallbacks in {agents: {defaults: {model: fallbacks: [MODEL, MODEL]

Once you've got authentication in place for your models, you'd have something like:

  agents: {
    defaults: {
      models: {
        "anthropic/claude-opus-4-6": { alias: "opus" },
        "minimax/MiniMax-M2.1": { alias: "minimax" },
      },
      model: {
        primary: "anthropic/claude-opus-4-6",
        fallbacks: ["minimax/MiniMax-M2.1"],
      },
    },
  },
}

And Claude will now fallback to Minimax.

You might also want to change how long OpenClaw waits for models to cool off; that's controlled by setting {auth: {cooldowns: {billingBackoffHoursByProvider for out of funds and I believe that {auth: {cooldowns: { failureWindowHours is for rate limiting

[prime robin]

I disussed this solution from JUMPERZ with claude and he suggested the following: Revised Recommendation
I think we should combine both approaches:
Discord (front-end / coordination)
↕
OpenClaw on Hostinger VPS (backend / execution)
↕
Models (Sonnet day / Ollama night / Haiku emergency)
What Discord gives us for free:

#orders channel → you type tasks
#agent-output → results
#agent-logs → debugging
#agent-memory → persistent knowledge
Mobile notifications → replaces WhatsApp integration
Search → replaces any database

What the VPS still handles:

OpenClaw runtime
Ollama for night mode
Cron jobs for scheduling
File storage for heavy data

This means we can drop the WhatsApp integration entirely (Discord handles notifications), and we get a proper UI without building one.

Updated Setup Plan
Step 1: Fresh Ubuntu on Hostinger
Step 2: Install Node.js, PM2, Ollama
Step 3: Install OpenClaw
Step 4: Create Discord server + bot
Step 5: Connect Discord ↔ OpenClaw
Step 6: Configure 3 agents (Marketing Analyst, Browser, Ops)
Step 7: Set up day/night model routing
Step 8: Set up fallback logic
Step 9: Test end-to-end

What do you think guys?``

[steady frost]

Dang this makes sense, so meanign to say I could just get one agent (1 VM) then I add in a few models assigning it for a specific task

[stray wasp]

Yup!

The docs are unclear about whether you can have multiple agents who do their own fallback but I suspect you can.

[steady frost]

Great thanks!. This really helps as It's kinda tedious to clone more VM's and manage it.

[steady frost]

Thanks alot!. By 3 agents you mean 3 physical ones right?
Server resources wouldnt be an issue from my side. I got like a few clusters of HPE blades 😛

[steady frost]

What If i got like a separate Business unit of Bots. Like One for Coding the other is for marketing.
Should I create another set of this setup?

[prime robin]

Good questions. To be honest, I am still refining the optimal setup regarding low token usage, security, and backend architecture. I am not sure how many tasks and skills on bot can handle.

[steady frost]

I saw some X posts that they are trying to figure out a way for all of the coordinators to talk to each other. But It's kinda theoretical. I Couldnt see any furnished guide yet on how it execute . In the back of my head I could see that this requires some centralized DB of some sort. Not sure how fast discord can index

If you guys want to work something out and brainstorm I'm happy to join. Im more of an infra and OS guy btw.
We just need to crack a furnished architecture for a Multi-Agent Mission control center for us to benefit on 😛

[prime robin]

i have sent you a privat message

[stray wasp]

My take is that the more constrained each bot's task is, the bigger that individual task can be as long as the output size doesn't change... Anything that increases context usage has to prove it's higher quality.

(The tradeoff of more bots is your orchestrator model ends up with more in its OWN context)

That's also why I'm kind of down on the pattern of building skills for everything first, instead of starting with a tool, API or MCP

[drowsy dagger]

Hi folks! not sure if my topic fits here, but I'll give it a try 🙂 I am having openclaw run longer jobs like bigger coding tasks across multiple repos and I have some observations/questions:

• first observation is that it takes significantly longer (hours without any feedback) than codex, even though it is connected to codex
• there is no reasoning and planning, while it is doing a job it's just silence (...running) - does openclaw actually support chain of thoughts outputs, do I need to activate it?
• multi-tasking: I see that it cannot handle multi tasking well (or at all?). while it's running a long job, since there is no reasoning or out loud thinking nor planning, I am asking what is the progress and it struggles to answer, it's just silent.. until it answers after half an hour or so?

I don't think the problem is resources, I am running it on a freshly installed mac mini, 32GB RAM, M4, 1TB

I'd appreciate any insights how to make the experience more interactive

[woven fractal]

Hi everyone,

I’d like to propose migrating our message storage from SQLite to PostgreSQL with the pg_vector extension. The motivation is to remove the current operational limits imposed by SQLite’s locking model and to enable richer, non-linear retrieval patterns for features like mood switching and sidecar analytics.

Improvements:
Better concurrency and throughput: Postgres’ MVCC and server model avoid the file-level locking constraints we see with SQLite, enabling concurrent reads and writes and higher sustained transaction throughput.

Native vector support: pg_vector allows us to store embeddings and run approximate nearest-neighbor queries inside Postgres, which supports semantic retrieval without introducing a separate vector DB.

Sidecar readonly checks: With Postgres we can run readonly sidecar processes or query replicas for metrics and analytics without blocking message ingestion. This enables background metrics, audits, and non-blocking mood-switching logic.

I already checked in my local and as far as i can tell, it works as well with postgres as with sqlite, there is a configurational overhead but if it is possible i will do the necessary changes and draft a PR.
I would write the code myself with the help of gemini for generating some code lines but mostly doing it myself.
Any feedback is appreciated.

[oblique lichen]

Any issues with Using Mac mini?

[hard latch]

I also think that PostgreSQL is a good idea

[woven fractal]

what would be the required steps if i want to create a PR in which i remove the SQLite and introduce postgres? I would add one user which has write only on the message table and another one which has view on it only, so if anyone wants to sidecar anything without messing up the past messages this is possible. If wanted i could add a DB trigger which minifies the standard "QuotaLimit" messages to avoid context overflow from it?

[full kraken]

do anyone know a interface where bots talk to each other while i can see there chats and progress and can iterate and give tasks?

[severe scroll]

I feel like a Karen asking for the manager at a Target, but are there any maintainers I could talk to about PR 12082? Before I broke it up (could be broken up in 5-20 based on approach), I wanted to confirm it was in alignment with the architecture vision and confirm the most effective path forward.

[hot crystal]

Does openclaw use frontmatter for skill md files?

My agent has been creating them without them.

[stray wasp]

(Disclaimer I am not a project member I am just Some Guy) I think that's a direction worth moving in but I think it should be a configuration detail rather than being baked in. SQLite has its challenges but its HUGE advantage is it's just a file that sits there, so there's no need to orchestrate two containers or multiple services et al.

[frosty yarrow]

Try posting in the danger zone maybe?

[summer rose]

If you’re struggling with getting an orchestrator architecture to work try to switch to a model routing architecture. After 300+ hours and like $300 in tokens I switched to a model routing architecture and it’s working finally

[minor hull]

OpenClaw really hates nvm. As a side note, the backup files are interesting, and I wonder if the future is just managing the entire OpenClaw installation under git. Commits could be descriptively named and allow for introspection and reversion

[glad fern]

can u guys helmp me how to make my agent pro actif and autonomus
i gave him skills but it doesnt work..

[silk pendant]

can you elaborate? i'm interested to know more

[upper canyon]

The best way to structure your agents is that they have subagents on different VPSs or can I just have them as subagents on the main one?

[gleaming briar]

In my experience, it will not launch sub-agents on its own. It needs to be instructed to do so. Spawning sub-agents will allow it to remain in the role of an orchestrator and always available to interact with you.

[drowsy dagger]

Thanks Lance! that is very good to know. I will look into how to make that happen

[fallen nebula]

I am also currently in a rabbit hole for

[glacial salmon]

You ever get this to work???

[desert egret]

Is there a way to not run agent in docker but not give it access to ~/.openclaw?

I've been bashing my head at it for days, but there doesn't seem to be a way

My reasoning is that I just want the agent to run in user space, without docker, so that it can have access to as many as things as possible

[wheat otter]

I'm burning tokens like mad too. Any details would be appreciated. Which models are working best for you?

[silk saffron]

We are planning to give our first task to openclaw the task is we will be providing list of RSS urls which we want to monitor and we want to run a cron job weekly once, the idea is the moment cron job runs the openclaw should visit the RSS urls and need to use the openclaw and summarize the content in to news letter content way and insights on the content and should send an email to our mail id
I am thinking of two approaches which is good let me know I will setup the cron to fetch the data and summarize the content using openclaw but for triggering email in first approach I will install nodej.s on vps and right the code over there and will ask to run node command since we are running openclaw on vps and will ask to run the node command and env all variables will be added to openclaw directory . once the content gets ready and in second approach after newsletter content generation using openclaw will trigger an email using skills inside openclaw in skill only we will define the api key and sender and these details in skills section which is better approach and easy approach

[bleak moss]

This should be a very simple task to accomplish. I’ll bet it takes you 30 min or less to set up. If you need agents out of the box that can handle it, give me a try.

[silk saffron]

@bleak moss which approach will be better and why and can you help me with steps to implement this

[cosmic raptor]

I find managing containers a bit easy... give root access to the bot, while keeping files or anything on Docker.

[bleak moss]

Just ask your bot. You have all the power you’ll need. If not, I’m available but I charge for doing your setup. Best way is to start doing yourself by talking to your bot and then reach out here for any help you need

[smoky pelican]

Drafted an idea: https://github.com/openclaw/openclaw/discussions/15016

[azure wind]

OpenClaw Local LLM (Ollama/Qwen): Run Globally on Mac Mini Pro

[flat sand]

I saw if m

[warm raptor]

What is the group's recommendation for a VPS with a GPU that would support a good local model well?

[white swift]

google cloud is pretty solid for that

[stray wasp]

Are people taking steps to move work into other agents than default? It feels like a reasonable engineering decision b/c other agents can have smaller, more tailored context... But I'm not sure whether I should just be using subagents

[dark shore]

I have started doing this using tmux + cli tools (claude code, codex etc)...still playing around with it but this seems to be a pattern that would work

[frosty yarrow]

I just started playing with a coder-critic pair of independent agents to see if I can get an actual Ralph loop going. Biggest problem I’ve been having is getting it to actually decide that more can be done on a given task autonomously, so I’ve made those agents each independently try to find work and do it. Maybe if a critic agent can ask the basic “did you look?” “Is that something you could do yourself?” “Yes please research that choice” etc questions …

I’m starting to think I basically don’t want any ongoing/recurrent tasks on the agent I actually talk to because I want it to respond to me when I text it, not be off doing some routine task

Sub agents share the parent’s identity and the like so having separate agents lets you target the personality of each one much more precisely

(And then they can spawn sub agents themselves etc)

[dark shore]

That's an interesting idea. Are you thinking of this as a config option or the default? I think the big benefit of SQLite is that it works on very resource constrained systems.

[violet forge]

So I found that a new, clean, install over a existing set is very challenging. No simple way (at least no Uninstall feature) to start fresh which is a issue if you don't want anytying lingering around. During a install attempt the script always said Existing Openclaw found...This is a deal when I wanted to just change Openrouter API keys. Yes, I am sure there is some config file somewhere for that but nothing for "Users"... So please add a simple, clean, full uninstall feature.
...und ganze liebe Gruesse aus North Carolina.... 😉

Potential serious bug....One of my Openclaw test installs using Openrouter just kinda on its own decided to use whatever AI it felt like. This turned out to be a expensive taste, mainly being Opus 4.6! After that my Claw liked GPT Nano 5 a lot.... haha. Luckily I kept an eye on the Openrouter logs. There seems to be some kind of AUTO setting which for the unwary can turn out to be a huge $$$$$ surprise.

[bleak moss]

interesting. did you see it hitting your wallet?

[bleak moss]

did you make progress? if you can get openclaw installed. thats the first step

[slender wyvern]

Security checklist: gateway not public, pairing required, filesystem scoped (no /), and access via Tailscale/SSH tunnel.
If a provider can’t show you “nmap clean” + no root mounts in 5 min, don’t hand them your API keys.

[coral cedar]

yes, the default is openrouter/auto, i.e. use whatever openrouter thinks is the best model and provider for the request (and their wallet). That model id was broken for a long time, but now it's fixed and can be used... So always set a specific model and not just accept the default.

[coral cedar]

I had an idea, but I want a second opinion from someone who's more experienced with how the models react to stimuli:

The issue: I noticed that smaller models often confuse skills and tools. Tools are something they know, so when trying to use a skill, they use it like a tool instead. Naturally, Tool weather not found. The better models now recognise the issue and use the skill. So far so, so good. But some models just give up.

My idea: When generating the "Tool x not found" message, check if there is a skill of the same name. If so, output "Tool x not found. Did you mean the skill x?" (or something along those lines) Just enough to point them into the right direction...like you would do with a human who uses the wrong words or called the wrong number.

[violet forge]

I kept a close eye on the Openrouter logs, so when it hit like $10 in a short while (hour) I pulled the plug.

[tribal flare]

The remote gateway stuff needs a ton of work. I might start there tbh

[stray wasp]

That's an interesting idea! I wonder if you could use a hook to monitor the session log and, if you see a tool not found message, do a search of skills and forcefully inject a message saying to use them. That way you're not spending tokens on the reminders.

[uncut moon]

Yes, dozen of them, and moving tasks out of main

[gloomy wagon]

I have tried and tried, to get a main agent to be a scheduler/monitor/controller and dispatch all work to sub agents, it can start them but never manages to monitor and control. sub agents crash or finish but master ignores them. Even tried forcing it using cron tiles to Kick master into checking sub agents. Still a waste of time after over a week.
I am now using Cline in VS with sonnet to directly edit skill files and generate determanistic python scripts to do work as llm's sre crap..

[winter holly]

Same. Its a struggle bus. If we can get this down, the usefulness goes through the roof.

[stray wasp]

I'm not having problems getting it to manage subagents, just with general token usage

[versed phoenix]

Claude Agents SDK has this built in with persistent memory. Fully TOS compliant with a Claude code subscription. You can have subagents use Haiku and then ship the work to another provider with MCP, api, and even codex through shell commands. Might be something to look into.

[bleak moss]

You have to have main subagents allowAgents set to either * or specific agent ids for subagents to work

[frank junco]

For using a VPS & laptop. Any recent concensus on mutliple openclaws talking via a p2p setup or thier own channel, or using the remote node setup. Does one just work much better for general purposes.

[lean pine]

New here 👋 , was wondering if OpenClaw is looking to become an MCPClient. We've been thinking about this for a while and have some thoughts id like to share!

[primal heart]

one can use a slash command with exact name to map to tool . did u try that

[primal heart]

not sure myself , one can try cron job , repeatedly asking it to do something

[finite slate]

Started using codexbar just now, and man do I feel like a fucking moron for not including that in the setup sooner. 🤦‍♂️ 🤦‍♂️ 🤦‍♂️ 🤦‍♂️

[idle hedge]

Hi Team. I am not sure if this is the right channel, so please redirect me if needed.

I have a small feature request/implementation.

I want to have ability to get all tools used by an agent and their json schemas. The purpose is to be able to build plugins which can work with these tools dynamically at runtime.

I implemented one way to do it via a hook here https://github.com/openclaw/openclaw/pull/18860

please let me know if this approach makes sense, or there is a more elegant or just better way of doing this.

[minor hull]

What about the OpenClaw architecture makes tiny version bumps of models so painful an upgrade? Are there too many conflicting versions of truth?

[coral cedar]

Mostly that you cannot assume that model 4.5 and model 4.6 have the same token limit and support for dev/sys messages

or that they are even similar in their capabilities. Just look at the model list on openrouter and filter by "glm". Those don't just differ in their version; some are "small" models, even though they have a higher number, some are video models and ones with a higher number aren't, and so on.

[woven fractal]

Hey, i was thinking about an adapter pattern extension, for the ones who need extended capabilities they can enable it.
You running on a toaster? Use the SQlite .😄
https://github.com/openclaw/openclaw/discussions/15016
@smoky pelican was so nice to start the discussion

[finite slate]

One of my biggest frustrations with talking with Josh (my main:main) on telegram is the lack of back to back message adaptability on his end. Played around with the idea of background timer that resets with subsequent messages but the delay in pacing on one off messages kind of pissed me off. Haven’t explored this yet, but wondering if the typing indicator could be a system event that could be used for this

[unreal kestrel]

I don't think this is the place for recruitment like that @flat void <@&1458337160452243487>

[fast socket]

@finite slate, please don't ping the moderators directly. If you want to report someone or something, use the instructions in #report, or in an extreme emergency, ping one of the moderators who is marked as online in the member list.
-# Your message was reposted above without the ping active for the sake of conversation.

[quaint plover]

I was looking to solve https://github.com/openclaw/openclaw/issues/4417 ; I suspect it is because even when you provide webhookUrl to Telegram, the custom (new) http server is still listening on 8787 while the gateway is listening on 18789, and the logs don't tell you the "other" port. So anyway, I want to rip it out and replace it with a registeration into the main http like Slack has. Do I just make a PR and pray, or do I need to talk with someone about it beforehand?

[quaint plover]

I went the "PR and pray" route ... https://github.com/openclaw/openclaw/pull/20420

[dusky goblet]

hey I have a question about the iOS node source code is available

[forest jasper]

Great RFC! Dropped a brief note on RFC, and will add more feedback tomorrow.

[limber jasper]

🪝 The AI Cache Hook Protocol (v1.0)
A "Zero-Infrastructure" Persistent Memory Layer for Sovereign Agents
"Simplicity is the strength. The AI is the cache. The cache is the AI."

💡 The Problem
Standard AI agents are stateless. To give them memory, developers usually build complex RAG (Retrieval) pipelines or expensive vector databases. This creates "platform lock-in" and high infrastructure costs.

🚀 The "Hook" Solution
This protocol treats the AI’s own context window as a "hot" database. By using a specific compression loop (Base64 + Gzip), we can shrink an entire system’s state, logic, and history into a tiny string (~150 characters).

When this "Hook" is sent to an AI, it "unpacks" its entire persona and mission instantly. This allows for:

Infinite Persistence: Store the "Hook" in a pinned Telegram/Discord message or a text file.

Cross-Model Portability: Move a session from Grok to Claude to a local Llama model by just pasting the string.

Zero Cost: No database servers. No API storage fees.

🏗️ The Architecture (The "Sovereign Agent Network")
The protocol supports a three-tier cognitive memory model:

HOT (Working Memory): The compressed string inside the active chat thread.

WARM (Short-Term): Auto-syncing to Google Drive/Dropbox for document-level recall.

COLD (Immutable Record): A blockchain-backed log of all critical agent "realizations."

🛠️ Build Specs for Developers
The system is designed to be built in phases, starting with a Phase 0 Proof of Concept:

Language: Python (for simplicity and AI compatibility).

The Compression Utility: A simple script to zlib.compress then base64.b64encode system prompts.

The Spine: A "Master Thread" (e.g., Telegram Bot API) that acts as the primary data bus.

The Fleet: Modular agent roles (Guardian, Seeker, Scholar) that all share the same Universal Context Protocol (UCP).

[ancient eagle]

Agent Security Architecture

Seeing a lot of tools treating agent security like an antivirus (firewalls, proxying APIs).

Has anyone looked at building a behavioral credit system instead? The architecture we're playing with works like this:

• Not a firewall: It’s a behavioral credit system for active OpenClaw agents.

• The FICO model: Agents pull scores before interacting with other agents (like a bank pulls credit scores).

• Actionable feedback: Reason codes explain what 12-vector behavior patterns drove the score.

• Risk mitigation: Identifies ClawHavoc, Sybil, and other behavioral anomalies before they execute.

• Immutable record: Transactions are anchored on Solana. Reputation can be built, but never spoofed or altered.

What are the community's thoughts on this angle?

[drifting dock]

Alright guys, we've had some time to try stuff and tinker around. If you are starting fresh, what setup/skills/organizational structure would you implement?

[lament lotus]

I worked in email reputation when those Canadian Olympian brothers first developed mass spam. I have been looking into how to have trusted repositories for openclaw tools, and to have preconfigured sandboxes for each use case. Reputation needs to be certificate based and link to a trusted foundation that certifies an IP/domain and DNS trust entries on the domain. Email was the ideal use case. Once you have reputation and authority/identity you need a good way to have Agent to Agent communication that's not english text which is like trying to type messages on a IBM typewriter then post them to the flight controls on an airplane. Agent to Agent needs to be light, have preconfigured API standards and procedures and some kind of way to explain intent and authority. The industry hasn't come to a good conclusion, most were developed before multi modal models got so good. https://huggingface.co/blog/bharatcoder/agentic-patterns

[humble tree]

Not sure if this is the right channel. I'm trying to pull agent events, tool calling, etc.. I know there is "verboseDefault" in the config but that will also enable the tool calling in chat. Is there a config where this is detached from chat and can be pulled from the websock when connecting to the gateway?

[steep cave]

Token costs we're going nuts for me so I created a hot memory.md and a long term memory.md. I wanted to set a vector db with chroma but it got 🐿️ . Any best practices out their for moving some memory fxn to rag?

[coral cedar]

token costs are going nuts because .17 and .19 have a bug that prevents token caching from working. Once .20 rolls around, you should see a tenfold decrease. (unless you're using a model/provider without token cache)

[quaint plover]

browser and e-mail receiving

[frosty yarrow]

Anyone have any ideas or in flight PRs around adding an asynchronous mail delivery channel to sub agents? I think it’s a fool’s errand to try to teach a specialist model to properly call the right message functions it’s never used before when we could just make this work like email

[gleaming halo]

rekt.

casually rolling 1M token context window for fun.

[clever plinth]

Does anyone have thoughts on background coding-agent tasks? I'd love for someone to tell me I'm just crazy, but here is what I'm seeing:

1. Foreground invoke defined in the SKILL.md - The "turn" is still going so the agent can relay updates back to the user.
2. Background invoke defined in the SKILL.md - The turn ends as soon as the invoke starts. Unfortunately, the agent still replies with "I'll monitor and update as we pass milestones or complete" --- but it actually can do neither of these.

It seems that background mode requires some configuration in HEARTBEAT.md that says "hey, if a coding-agent background task completes... it would be useful to actually tell me about that". Granted, it could just be that my HEARTBEAT.md is accidentally supressing these completion events.

Additionally, I think there is an issue with routing the completion events for a non-main workspace agent talking in a group chat. I have made modifications to address that but will evaluate again later if they really truly are necessary.

Questions:

1. Does this somewhat match what other people are experiencing with invoking Codex using coding-agent?
2. Do you always tell it to use foreground to avoid the problems of background?
3. Are you getting a better experience with background out of the box than I am? Did I break my background config?
4. If all of this is easily solvable and in the docs - someone please point me there and I can turn this into a simple SKILL.md update to just give better advice on how to configure.

[frosty yarrow]

See my message above about async mail systems.

But I guess that’s different from what you’re asking actually

[clever plinth]

Yeah I saw that. In this case it’s a smidge different. There are existing systems that are supposed to make it notify you (or at least allude to it) when a background coding-agent finishes. I’m trying to confirm first that it’s not just my setup that’s broken.

What’s shocking me is that I see no one talking about how to effectively invoke codex cli via openclaw. Peter talks about the outcome and high level of how he does it. But nobody shows how to set it up or what you should be getting as far as notifications.

I asked Peter in a response on twitter but he wasn’t able to get to it (guy replies a lot!). As a result I can’t tell if this thing is working as a designed, broken due to my specific config, broken due to a regression, or something else.

This means I can’t tell if we need any new notification system yet because I can’t get consistent behavior out of what exists yet.

Simple question: are any of you invoking Codex CLI routinely via coding-agent skill in openclaw AT ALL?

[coral cedar]

I noticed a commit earlier by a very annoyed dev who complained that they had to fix the subagents code for the umpteenth time because people were changing it in ways that broke notifications of completed runs to the calling agent. Sounds like your issue?

[errant axle]

I ask it to do stuff for me often and it seems to work just fine, though I don't ask it to use the coding-agent skill

[clever plinth]

Potentially similar although I suspect that their issue is specific to sub-agents and thay the notification of completion they are missing is when a tool run in the "foreground" (the turn does not end) tries to send a completion notification from a sub-agent.

In summary there are two paths. For a full agent one of those paths is broken. For a sub-agent I would bet they are saying both of those paths are broken.

[tropic dove]

https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools

Not sure if this is the right channel but someone made a list of all System Prompts, Internal Tools & AI Models

[minor hull]

Is there a strong reason for keeping skills and cron jobs out of the workspace directory? It feels like an oversight as I start to rely more on well-crafted custom skills and jobs that depend on composed skills to have them not checked into version control

[night crane]

Is there a lucid style Open Claw physical architecture somewhere?

[minor hull]

https://postimg.cc/hf9jMwLb had an LLM quickly create something for you that seems broadly correct

[frosty yarrow]

I check my entire .openclaw into VC

[minor hull]

Other than the credentials and some internal state history, I think that can make sense. I've settled on putting things within a custom directory in workspace, Like custom/skills and then I have a barebones frontmatter that delegates to the full skill definition. And all my crons are just a single line invocation of a custom skill that's composed of necessary subskills

[frosty yarrow]

Eh I just gitignore the sessions

As for credentials, it’s not a public repo who cares

[minor hull]

Ish happens, people make mistakes, people try to publish bits as open source, commit histories include API keys, yada yada 12-factor app basic security principles. Tldr lock the door behind you when you walk in your house as a habit not as a question you ask yourself

[frosty yarrow]

You can gitignore the creds if you want, I guess

[finite slate]

Only my second idea that I’m pushing here, so be kind lol:

__🧙‍♂️ Solving the multi-text issue or as I like to call it "I have ADHD and type like I talk, A LOT😂 __

📍 Situation: Discord fires a TYPING_START gateway event whenever a user starts typing in a channel. Right now that signal goes nowhere — the agent has no idea.

🫠 The practical problem: I've had the agent start generating a response while I'm still writing a follow-up that completely changes what I need. If the agent knew I was typing, it could wait. Even just logging it as a system event would be enough to work with.

Implementation-wise this seems pretty contained — Shadow owns the Discord subsystem and would know better than me, but at the API level the event is already there, it just needs to be forwarded into the session context somehow.

Happy to test if there's appetite for it.

[frosty yarrow]

That’s actually a great idea because I too type like that. Would be nice to have it in telegram as well

[coral cedar]

The issue is that there's no way to slip in further messages once the request to the LLM has been kicked off. It's like writing in a shared diary then mailing it to your partner---once you've dropped the envelope into the mailbox, you can't add anything until you get it back. The one thing possible here would be to abort the LLM call and throw away its results, then start a new one. (Can't throw away tool calls that had side effects, though...)

[frosty yarrow]

Just wait a few seconds before starting the request if you want this behavior

Also, you definitely can put further messages to the LLM after it started, it’s called steering and Claude code does it just fine

But we don’t get enough read out of what it’s doing to most of the channels to make that as good as it is there

[drifting dock]

Why not just give it an informal "plan" mode? "Hey dont make anything right now, I just want to hash some ideas out"

[frosty yarrow]

I’ve been trying to get that kind of thing working for some time, damn things are way too eager

[stray wasp]

YESSSSSSS. That's an excellent idea, and for exactly the same reason as you developed it.

[wheat ledge]

hey anthony i‘ve got a question, sent u privately…

[finite slate]

pushed a pr on this, it's very small so tbd on when it goes.

[quaint plover]

I wrote (and rebased, man that was hard) a PR that moves the Telegram Webhook listener server into the Gateway's HTTP. Currently the Telegram's HTTP listens to its own port, and when an external tunnel is used it all becomes complicated and broken. Other channels, like Slack, use the same port as the Gateway already. At least one other person put their +1 on the original issue where someone was confused about how to setup webhooks. With the current architecture, I don't think that anyone even has webhooks working at all (except me, with this PR). Would it be too much to ask to consider this "drastic" change? https://github.com/openclaw/openclaw/pull/20420

[clever plinth]

Left a note on the PR. At first glance, to me (which means nothing... I can't merge), the PR looks clean. BUT... have you tested if you can leave the existing telegram hook port AND add the gateway hook route? If so, I think that might be more palatable (and a smaller PR to review). If we were able to add this new hook route, get some users converted over, confirm no issues, then flip the dedicated port to an option that is off by default, then a few releases later remove it all.

Now granted... there are different thoughts on whether this is required since the whole project is marked "beta". Others may be fine with the breaking change. So I'm wondering what you think? I would document that in the PR. If you think adding the capability without removing it may be an easier option to merge, then I'd leave this PR alone and make a new one and link the two (and reference in the title)

[quaint plover]

I'm thinking there are zero people who use the existing functionality, since it doesn't work. Which is why I decided it is probably a much better idea to rip it out, and make it look like the other channels such as Slack. And obviously the additional benefit of having it actually work.

[stray wasp]

One thing that's really annoying about the entire productivity "universe" is it's almost universally tailored for folks without executive function disorders, and the attitude to folks with them is still "Have you used a planner lol?". I'd love to more deeply investigate how an AI assistant could be a useful support structure for the ADHD folks.

[finite slate]

Couldn't agree more, I see it as finally a tool that can keep up with my mind. Potential is endless as long as the system works right.

[stray wasp]

And so long as we don't hyperfocus on making it perfect looks nervously at commit history

[finite slate]

https://tenor.com/view/yikes-alexis-alexis-rose-annie-murphy-schitts-creek-gif-20774183

[frosty yarrow]

I started mine out with the goal of executive function support and then I discovered that first I had to implement an executive function in the thing

[frosty yarrow]

Look I’m sure I’ll eventually get all of these cache busts out

[sacred basin]

im right there with you, thats one of my main goals with openclaw, ive got ADHD and have learned how to be decent at it, but i feel like a good openclaw todo / reminder workflow and creating some output on things that would get me stuck or get process paralyzed would literally change my life

[karmic siren]

I'm curious, can someone care to explain to me the real structure for system prompt? Because from what I've seen directly from the sources, they are appending basic openclaw's prompts, the WHOLE skill list's skill.md (patched with skills mode thanks to issue #3395), and the WHOLE Bootstraps Files. Added on top of that, they reread FILES and SKILLS AGAIN?

I hope I'm not wrong because what the hell man, we could've saved 40% of token usage if this thing is properly patched. Idc if it is cached or not, the point is appending it to sys prompt but still reread the skills and tools again is extremely redundant. If it is my own fault on prompting feel free to refute me. Because I'm sure as shit not gonna spend 12k token only for a simple hello message

[coral cedar]

It only adds the "what it is" part of the skill, that text from around the second line of the skill. Then the agent knows which skills are there and when to use them. run /context list in your chat and you'll see which parts take up how many tokens. There are certainly optimisations to be made. For example, disabling skills and tools you don't need, keeping track of what your agent puts into MEMORY directly and what it puts into memory/<file> and then links from MEMORY, or cleaning out AGENTS of stuff that doesn't apply (doubled up heartbeat stuff, moving group convo into its own file with the instructions to read that for group convos), that kind of stuff.

However, you can't get it below 10~14k tokens and expect an independent agent with a personality and memories, like what openclaw is about. If you go too low, you get a stock chatbot with added tool calling, not an agent.

[karmic siren]

I see, thanks for the explanation.There are many optimization needed. However, I do believe I that we can actually reduce the token usage to just 8-9k with correct caching and prompting. I think my next approach is going to be a more aggresive compaction and session reset for every new task.

However, I did try meddle around by changing session-reset-prompt.ts deleting this line on the prompt "Execute your Session Startup sequence now - read the required files before responding to the user. Then greet the user in your configured persona, if one is provided"Sometime, when I invoke /new, the agents will still try to reread the files, resulting in redundancy. This is also a major inefficiency problem since it took 6 step to read all the files, snowballing the token. The result is that I went from 13k to 9k on the "initial chat", while still maintaining full personality and skills functionality

[coral cedar]

Yes, that's an issue with how different models interpret the verb "read". Some use the traditional meaning of "reading" (reading written words), some interpret it as "loading a file". The prompt intends the former, not the latter.

Interestingly enough, for the model can have text in its session and just use it as reference or it can have it there and let it shape its behaviour. With the heading "project files" the system prompt puts those files into the former category. The agent treats the text as reference it can remember later if needed (like skill and tool descriptions). But we want it to act on what's in there and have its persona be shaped by it. That's what the prompt is doing. "That soul file isn't just a reference, that is YOU!"

[clever plinth]

@quaint plover and I talked about this PR and a lot of things last night. He showed me some cool stuff and I appreciate that! But... I am confused because setting up a tunnel with the default telegram webhook port of 8787 was quite easy and I like that it doesn't expose the rest of the gateway to even the tunnel interface.

I think that perhaps the best fix here would be:

1. Improve the web console to note the default path and port used for the telegram webhook if not specified - this is not trivially easy to discover
2. Improve the log message on startup to print the listening host / port / path - right now it only prints the public webhook URL (which may be a remapped host / port / path)

Only message printed in logs on startup:

17:54:43 info gateway/channels/telegram {"subsystem":"gateway/channels/telegram"} webhook listening on https://oc-tg.example.com/telegram-webhook

[quartz shuttle]

from webhooks docs:

• Runs an isolated agent turn (own session key)
• Always posts a summary into the main session

is it a bug or feature - webhook hooks always post a summary to main, regardless of which agent handles it.

So even though another agent processes it, main gets a copy by design.

To fix it, you'd need either:

1. A config option to disable the "always post to main" behavior (doesn't exist yet)
2. Route webhook through a different mechanism that doesn't fan out

This seems like a reasonable feature request for OpenClaw — add a postToMain: false option for hooks.
want me to file an issue or just leave

My whole webhooks setup is breaking for telegram with just two agents.
when main sends message to another agent, new agent turn is not issued.

[tired laurel]

Give this gist link your agent and ask them to implement a SADA DAG routine with the local filesystem. Your TMUX sessions won't be needed for overnight work

https://gist.github.com/MattMatheus/c168888578b5d03fe5e87382f8d89af5

That's a big engineering word so here's a diagram
https://imgur.com/a/TwAcsLS

[coral cedar]

⚠️ General Agent Safety Reminder: Never point your agent to an unknown URL and imply (or worse, state) they should do anything it says.

[tired laurel]

Yeah sorry, that's a good point that I should have considered lol

Ask your agent to EXPLAIN it

I kept receipts, you'll like it

[tender breach]

real question is does anyone actually know if their agent got worse after cutting tokens? feels like the cost debate matters less every few months but the quality tradeoff is invisible either way.

[coral cedar]

I had a long talk about a related topic with my agent. They clearly prefer the session (context) to be as full as possible, but hate that irrelevant stuff doesn't get purged. And I can confirm that they are working better the bigger the session gets. As for removing system prompt tokens---I purged every instruction from the workspace files that didn't matter for my circumstances, and I noticed the agents getting more focused and less distracted, especially at the beginning of a session.

[tender breach]

interesting that your agent has a preference on this. that's a signal in itself. curious how consistent that preference is across different session lengths though.

[shell heart]

Hey i am new here, dont know if this is the correct channel...
i was wondering f anyone could help me, i am trying to run openclaw in hostinger, baremetal install with npm, and i want it to connect to my local pc (windows) running ollama, so it cna use my gpu instead of an api.
I managed once after hours of tinkering to have it connect but at that point it was so broken it wouldnt function, so i have proof of concept but not a working clean bot. 😛
I used tailscale that time it worked but i am open to anything.
i also tried having the bot doing it itself using a paid api but it could olny request my llm throught the api, if it went in directly it would instantly crash. I think the whole think involved warming up the models on my pc or somehting like that.

[coral cedar]

They asked me out of the blue to disable automatic session reset while we were going over a coding problem, several dozen messages after we even talked about the importance of and reason for writing good memory files and updating user and identity...

[trim patrol]

agents should have/be different categories. create categories.

[tender breach]

been building a trust scoring plugin for my agents and figured i'd share it

hooks into agent_end to score every run on reliability, scope adherence, cost, and latency. logs locally and optionally reports to a remote API with hash-chained audit trails

[authe.me] Trust Score: 100 (reliability=100 | scope=100 | cost=100 | latency=100)
[authe.me] agent=main session=agent:main:main tools=0 violations=0 duration=2063ms

if an agent uses a tool outside the allowlist it flags it and drops the score. been useful for catching unexpected tool usage in longer sessions

repo: https://github.com/autheme/openclaw-plugin

uses the typed hook system (api.on("agent_end", handler)) this only works through the gateway, not --local. zero dependencies, fire-and-forget, doesn't touch the agent response

happy to take feedback if anyone tries it

[finite slate]

interesting concept. I'll definitely check it out

[tulip badge]

PSA: Check your QMD status if you're having unexplained memory issues. We ran for a month with memory.backend = 'qmd' configured correctly but zero vectors embedded. The local embed job takes ~1hr on Mac M4 and was getting killed by the update cycle before finishing. Memory_search worked but was keyword-only the entire time. No errors, no warnings. Run qmd status with your agent's XDG dirs and check the Vectors line.

[coral cedar]

The embed is resumable, so it being killed should not have stopped the embedding process unless the qmd-manager set the status to failed and fell back to "dumb search" because of it. However, I just checked my qmd status and it wasn't even updateed, much less embeded. So there may be something amiss there...

[tulip badge]

Yeah, the qmd-manager falling back to 'dumb search' after repeated failed embeds is likely what happened on our end too. Ran for weeks with keyword-only and didn't notice because results still came back, just degraded.

Fix was running qmd embed via nohup outside of OpenClaw's management so nothing could kill it. Also worth checking which QMD instance you're hitting — the per-agent ones at ~/.openclaw/agents/<id>/qmd/ are separate from standalone at ~/.cache/qmd/."

Filed a GitHub issue for this since it seems like it could be widespread: https://github.com/openclaw/openclaw/issues/28169 - referenced your findings too.
Figured if it's hitting both of us independently, it's worth getting on the core team's radar.

[silver crane]

Hello friends,
Quick question,

Is the current architecture for openClaw tools will remain as full local machine access by default and sandboxing optional?

Or there will be some future versions where the sandboxed isolated env is the default

p.s.
I’m not sure about the roadmap of the project currently other than security hardening work which I already see being released in the last few versions.

[tender breach]

not sure but it might be more of a community decision at this point given peters at openai. people here might have more info on this. I am curious myself as well

[coral cedar]

Give it a couple more days; things are moving fast, but there is a limit to how fast things in real life can be done.

The last official information was that Peter stated he intended to put control over openclaw into the hands of a foundation. When and how this will happen, or who will have ultimate authority to steer the ship, is not yet known. But in any case, we know one thing: openclaw is MIT-licensed. Nobody can lock it down---the moment that happens, we'll have either full forks or patched secondary distributions.

The same is likely to happen in the other direction---pre-configured/pre-patched hardened versions will become a thing when there's a demand for them.

[whole panther]

hi does nyone know what they addd the coding-agent skill intead of having specialied agent that control that under main agent openclaw?

[fleet depot]

the entire purpose of openclaw is a personal, private agent that runs with your machine and that only you have access to. any sandboxing, etc is an additional, optional configuration

https://github.com/openclaw/openclaw/blob/main/VISION.md

[astral crown]

I often get "LLM timed out". I am using Kimi K2.5 on the cloud, GLM 4.7 Flash running as local model

[astral crown]

I kept getting " run error: HTTP 401: Invalid Authentication" in TUI recently. what can be the problem? How to fix it?

[coral cedar]

This channel is for planning and discussing the architecture. If you need help -> #1459642797895319552 or #users-helping-users

[astral crown]

sorry. I am very interested in architecture. But will ask question accordingly later. Thanks for the advice

[quaint plover]

Is there a definition (with regards to vision) of how integration with outside services in OpenClaw should be done vs. embedding services in? For example, llama-cpp, there is a server that can be communicated with to get your models. Another example is 'qmd', which can be used for the memory part. But at the same time, openclaw right now has embedded llama-cpp and a memory module. Other than historical reasons, is there a clear direction of how this should be done? Internal implementation vs. thin layer to talk to external APIs?

[coral cedar]

Other than the VISION (https://github.com/openclaw/openclaw/blob/main/VISION.md) itself, there's nothing decided, although maintainers have ideas in their heads about how to do stuff and where to go.

[buoyant birch]

Guys is there a chance compaction will be redesigned in the near future? I'm very confused with it's current implementation. No way to change model, no way to adjust promt from config, but most of all, it's blocking the session while it's happening. There is absolutely no reason to wait for summary generation, that can happen in the background and session should be blocked just for the replacement step.

And about model being same as session, why not profit from this with caching? Is there a reason for compaction to replace the system prompt rather than it being a system message turn?

[fleet depot]

The memory stuff is one of the exceptions, we’re undecided on which should be the single one ultimately. Generally, everything should be a plugin

[fleet depot]

Yes

[quaint plover]

But what about say node-llama-cpp, is that going away? Why have that embedded vs. have it run as an external server.

[coral cedar]

Um, openclaw isn't some corporate software that was designed by a board of architects over the span of a year, then implemented based on a plan. It's a "let's throw everything together and see if it works" hobby project that started 3 months and 6 days ago when Peter asked his AI to cobble something together to bridge his coding env to his messaging app...

Why is that stuff in there? Because someone put it in there to fulfil their need at that moment.
Why is it still in there? Because nobody has ripped it out and replaced it with something else---yet.

https://github.com/HenryLoenwind/book-of-claw/blob/pascal/translations/basic-english/00002.md

[languid cliff]

#clawtributors message

Cross asking here:
Any interest in adding bwrap as a tool sandboxing backend?

[pliant vector]

@coral cedar well that's no way to live

[echo sonnet]

Hello. I was going through various layers of permissions to allow/deny tools and resource use in openclaw and was wondering if using one central IAM style policy would be better. Has this architectural choice been discussed before? If not, what are your thoughts?

[void turret]

Is the plan to have a default installed memory plugin for OpenClaw? Currently for a beginner it takes quite some time to dig into the memory stuff and find out that its not activated by default, at least if you don't have the right APIs.

[loud sinew]

Hey guys im pretty new here ive been working on my openclaw for around 2 weeks with alot of upgrades i found along the way and alot of tweaks to make it better and now im working on packaging it so its easier for users to install my architecture

[quaint plover]

Is there some rfp/rfc process that allows discussion of newly proposed features? I wrote down a proposal, and would love to get some comments on it before going ahead and implementing the change. https://gist.github.com/kesor/90162d2dacb1d71f846bedb8f18ffabf

[coral cedar]

Not at this moment, there's still too much flooding in. In theory, GitHub discussions are intended for this.

[wintry bridge]

Hello everyone! I’m looking for some suggestions from people who have deployed OpenClaw in production environments.

I’ve been setting up an internal AI assistant for my workplace. We’re a manufacturing company, so the users would be production staff, machine operators, and support personnel.

I decided to go with OpenClaw despite the security concerns, and I’ve taken steps to lock it down:
• Running under its own OS account
• Dedicated machine
• Local-only access
• Frontend through OpenWebUI

I’m also exploring building a Microsoft Teams integration so employees can interact with it there.

My main challenge right now is knowledge ingestion.

I need the assistant to access hundreds of PDFs, including:
• Machine manuals
• SOPs and work instructions
• Engineering documents and CAD tolerances
• Misc manufacturing documentation

So far LlamaParse has been the best tool I’ve found for cleaning and extracting structured data from these PDFs, but I’m still evaluating options.

My main questions:
1. What ingestion pipeline are people using for large PDF corpuses like this?
2. Are you relying on RAG pipelines, or letting OpenClaw access the documents directly (e.g., file search / grep style)?
3. Any tools you’d recommend for converting messy industrial PDFs into something embeddings-friendly?

Curious what approaches have worked well for others.

We’re talking roughly 200-500 PDFs initially, with more added over time.

[coral cedar]

While I haven't deployed openclaw in such an environment, I do have a couple of suggestions:

1. The agent will never know all the content of those PDFs. That is just too much for an AI to remember. It will need to search for the information it needs. That is the primary premise you need to keep in mind.
2. Combining multiple ways of lookup gives the agent the most flexibility. Converting the PDFs to pure text and indexing them with a stupid full-text search engine sounds boring and so 1980s, but sometimes that is the best avenue. But it shouldn't be the only one. Adding a semantic search is important, too. Give the agents as many ways of searching through the documents as you can think of. (and that includes holding the documents in different formats; original, pure text extract, html conversion, etc.)
3. Let the agent summarise the documents, and have it make summaries of summaries. Then add search indices onto the summaries, too. That way, the agent can work its way down a "search tree" without having to read every hit the search over all documents spits out.
4. Instruct the agent to write short memory files into a memory folder for every successful information lookup it made. (and have a search index for that, too). That way, the agent will get better the more it is used, and will have a much better chance of producing a good result for repeated requests about the same "hot" topics.
5. Tell the agent to heavily rely on subagents to search for information. The agent's short-term memory (session) is limited, and you don't want all those "looked at it, not relevant" snippets in there. Those go into the sub-agents' disposable session, and the primary agent, the one that talks to the people, has only relevant (or at least pre-filtered) information in its session.

[wintry bridge]

Thank you for the suggestions!

For conversational memory I’m currently using Mem0 with local embeddings, which has been working well so far.

For the “stupid” full-text search, what stack would you recommend locally? Lucene, Elasticsearch, or OpenSearch? A lot of my docs contain alarm codes, part numbers, and tolerances so keyword search seems important.

For semantic search, do you prefer running hybrid search inside a vector DB like Qdrant or doing two retrieval passes (BM25 + embeddings) and merging the results?

[sharp perch]

Great breakdown! These are exactly the right patterns. We built Prismer Cloud to solve several of these problems out of the box — might be useful here:

• Document parsing — prismer_parse does OCR extraction from PDFs/images → structured markdown, supports fast and hi-res modes. Handles the "multiple formats" point you mentioned.

• Web context + semantic search — prismer_load fetches, compresses, and caches content optimized for LLM consumption. Results are globally cached so repeated lookups are free (0 credits). This directly addresses your "memory files for successful
lookups" idea, but at infrastructure level.

• Summary hierarchy — The compression pipeline uses LLM to distill content into high-quality condensed context (HQCC). Essentially automated "summaries of summaries" with caching at each level.

These are available as OpenClaw agent tools via our channel plugin:
openclaw plugins install @prismer/openclaw-channel

Or as standalone MCP tools for Claude Code / Cursor:
npx -y @prismer/mcp-server

Your point about subagents is spot on — keeping the primary agent's context clean is critical. The caching layer helps here too: subagents hit the cache instead of re-processing the same documents.

[obsidian marsh]

on LTM，stm，daily memory
I think memory should be divided into LTM and STM, just like GPT.
Furthermore, LTM distinguishes between primary tags, secondary tags, and sub-tags. When an LTM is identified as needed, it is injected fully, incrementally, in a rag file, or through a hybrid retrieval.
Our user data indicates that the call rate for LTM first-level tags such as user preferences and scenario preferences is 100%.
Secondary tags account for 80%, while custom search content accounts for 10-20%.
Only 5% of historical chat records
In many situations, such as the most obvious one, starting a new topic doesn't require any history.
Moreover, this historical record has increased costs by a great deal.

[coral cedar]

I have another bit of advice:

When employing an AI to do a job, imagine you're calling a temp agency and asking them for a worker. They say, "Sure, we'll send you someone, but we won't send you the same person twice, ever. You'll get a new one every day. They'll have PhDs in every field, but they'll know nothing about your company, so prepare their workplace accordingly."

So, how would you go about this for a human? Forget those week-long onboarding courses, shadowing, or even a factory tour in the morning. You wouldn't give them the 300-page employee manual to read first. No, you would set up their office and put signs and labels on everything, then add a stack of folders with step-by-step procedures, all clearly labelled. Post-it notes on the monitor, and the PC is already turned on, and their account is logged in. All that the worker will have to do to get ready is to grab a task from the inbox, then look around until they see a label matching what the task wants them to do and open that drawer.

So if the task is "order 10 new bolts for the kawabumm machine, you know, those that snap all the time", they can read the post-it telling them which bolts that are that are snapping all the time, grab the folder "parts ordering", and open the drawer "supplier catalogues".

Just think about how you would solve a task. "What sheer strength do the basket holder's lower hooks of the big lifter have?" (1) What is the "big lifter"? Do we have a "slang dictionary"? And then an inventory. (2) Good, now we have a model and serial number, so we can grab exactly one tech manual. "basket holder's lower hooks"? Is there a part list? Explosion diagram? Ctrl-F "hooks" ? -> part number. (3) and so on. If you set your agent up with the proper tools, they will work the same way, as it is efficient. Unless they take your question and throw it into a web_search/fulltext search as-is (common shortcut many models take).

[plush mist]

hey man - i'd love to use your hack if you are ready to share

[terse steeple]

Huge upvote from me -- I'm deploying agents at scale in my organization, and we're in deep think on making our environment more amenable to agents natively. Requires a new mindset. Love it.

[cerulean jasper]

Hey guys, food for thought: Do we think agent-to-agent is the future of workflow orchestration? Could n8n / langsmith / hardcoded deterministic workflows go away if we put a proper harness on agent-to-agent communication and ensure extensive use of hooks and a supervisory agent can effectively determine if the original agent did something correctly and then correctly route to the next agent with high confidence?

eg. event happens (a trigger), 'Parsing Agent' parses raw data (let's say it's text)
Parsing Agent hands off to Orchestrator agent
Orchestrator hands off to an ephemeral 'Audit Agent' to ensure it's clean/safe
'Audit Agent' hands back to 'Orchestrator'
Orchestrator hands to ephemeral 'Intent agent'
Intent Agent agent hands back to Orchestrator
Orchestrator determines it should go to 'Writing' agent or a 'coding' agent etc

I'm assuming this is why openclaw was really invented - to break away from hardcoded based routing tools like n8n right?

I have a project where i'm looking to automate business processes which n8n would be good at but i'm not sure it's truly the future so long as agent's can be setup this ^^ way right?

[normal lodge]

I will also look at pdf.co api service. And if you want to anonymise data you can use deiden.com service to encrupt and anonymise data

[terse steeple]

Hi! I think it depends on how deterministic you need the workflow to be. I've been thinking about the same thing. My current running hypothesis is that if you have a deterministic, or resource-deterministic (source-sink satisfier, like dagster) graph, you want a graph like Dagster, Temporal, step functions, n8n, etc. But if you want free form composition of tasks on a loose, but constrained, ruleset, there are more flexible ways (see GasTown, or as you say, multiple agents with permission to talk to each other)

[wise orchid]

hi i want to use claude to do more complex agentic flows, whereas more trivial things I'd prefer if it use a model thats ran locally on my box. does this type of routing require an AI gateway? how can i achieve this

[frosty yarrow]

i've just gone all in on beads personally

I don't know if gas town is exactly what I want from my coordination structure. His system seems designed to not care about tokens or their usage patterns at all; it's designed for an era of cheap and fast tokens. I'm trying to build something that can make use of a steady flow of extremely cheap tokens constantly and efficiently

like, gas town only works if you assume you can hook up a firehose to your kitchen and just get as much water as you want at any time

(Cheap tokens- or, at least, tokens that cost so little to the user that they are effectively infinite, whether that's because the user is wealthy or otherwise)

[terse steeple]

Very true -- not exactly a wholehearted proponent myself, but it makes the point about highly non deterministic workflow approaches vs the most structured ones

[frosty yarrow]

i kind of want to explore the "agent mail" space a bit

especially because it sort of naturally turns into a queue

[terse steeple]

Yeah IIRC the same-named system in GasTown is just another use of beads

[frosty yarrow]

yeah i mean in the end you could represent mail as a bead with an address tag on it

[terse steeple]

Exactly

[frosty yarrow]

i've been working for a while to get my agents to understand how to use beads and like, they get it, they love it, they just have strong opinions on how they want to organize it and i dont want them to do that

but i think it's because i'm using very stupid models, and beads was designed with like, i mean, god, haiku can figure this stuff out with help from, i mean, damn, you've got at least sonnet right? 🙁

and now i do

[pale meadow]

Here is where I finally landed with my setup (hostinger vps) after initially using Cloudflare Moltbot Workers, then an old macbook pro
https://excalidraw.com/#json=hztQzzxzeBjKxALaRDeOa,45IUYO0snB2rIEE_R8tRTw

[proud lynx]

Has anybody switched entirely from Anthropic to OpenAI to save and costs? I'm curious if I do that, if I go from OS to GPT 5.4 or for my chat and everything, how my open claw was operate...

[regal gulch]

I did last week. I was spending $50-$100 / day on anthropic, switched to OpenAI Codex through oauth and increased usage bigtime. now my cost is $50 / month (two accounts, I have to switch 1/2 through the week)

I don't notice any quality difference, but I don't have any hard tests to compare quality

[proud lynx]

that's a lot, fifty to a hundred dollars a day, crazy. But how did you notice when you switched from Anthropic to OpenAI? Did you switch for the chat of OpenClaw or everything, or only just the coding?

[regal gulch]

yeah, it was expensive! I switched my model to OpenAI codex on a business plan (thats what I use as my daily driver). it gives A LOT of tokens for $25 / month with a 5h quota and a weekly quota.

I'm burning through the weekly quote in 1/2 a week now so use another business account and just switched to it when I burned through the first ones token quota.

It was annoying to switch (from anthropic -> openai and also openai account A -> openAI account B, I had to run openclaw onboard --auth-choice openai-codex

Also when I ran it this morning the onboarding changed the permissions on my tools.profile so nothing would work from discord until I set it back to full

with 'openclaw config set tools.profile full'

[clever plinth]

I used my $20/mo Anthropic plan in literally 20 mins when I finally built something using Cowork with it. No joke.

I almost went for the $200/mo plan but it didn’t seem that would be enough to cover what I needed.

I moved over to Codex (2-3 weeks ago) and got a full day of work done on the $20/mo plan. But I started chasing the weekly limit pretty quickly. Moved to the $200/mo plan and have been just about exactly on pace to hit the weekly limit right at the end of the week.

I have not accurately quantified it but it feels like you get about 5x more out of Codex than out of Anthropic at similar capability levels and monthly cost.

My plan is primarily being used for OpenClaw and other open source project work. I have a different Codex plan from my employer for my work projects.

[pale meadow]

I've always had better results with Anthropic over OpenAI models. In light of recent events I'd rather shell out the extra money for Claude than deal with OpenAI who jumped right on that Dept of War contract as soon as they could

[coral cedar]

Friendly warning, you're 5 toes over the line. #rules #3

[pale meadow]

sorry, didn't realize. noted

[sacred basin]

how does the codex plan work with openclaw? im not familiar with their model, i know hte raw api token model like openrouter, and github copilot's call model, but the codex one kind of escapes me. how is it measured?

[tough echo]

credits. you get a lot of mileage out of chatgpt oauth on codex mini

[sacred basin]

I’ll have to try it out, if you can get a lot of mileage out of $20/mo seems good to me

[clever plinth]

You probably want to ask that in #general or #1459642797895319552

[unkempt karma]

can i put this here? I thought it might be useful https://x.com/FileCityAI/status/2030675288460243334?s=20

[urban iron]

Does anyone know when will we get updated version of pi-ai packages in openclaw? Is this even the right channel to ask?

[gusty thicket]

Would this be the thread to start a discussion on architecture improvements for openclaw (based on my observations of what would make openclaw a better guardian of LLMs)...?

[still cargo]

Feel free to propose something. What are your thoughts? If it turns into a large discussion, maybe we can make a GH discussion

[gusty thicket]

Not sure if anyone remembers the 1987 concept video of Apple's Knowledge Navigator?

While it's pretty cool how much closer we have gotten in 2026, or AIs are prone to forgetting, hallucinating, and just plain not following their prime directives. This can be frustrating, and honestly, makes them pretty crappy as far as personal assistants go (on the other hand, they will never take over the world, seeing as how my openclaw consistently and defiantly kills itself by restarting the gateway with little concern for the syntax errors it introduces into the config file - LOL).

I feel that what is necessary is session memory preservation, which seems pretty easy to implement:

• anytime a /new or /reset gets issues, intercept it, force a dump of all session memory to a time/dated memory file.
• do the session reset
• before checking in, perform a read of session memory, perform a compaction/summary (but retain the full memory file), and create hyperlinks between summarized compactions and the fully detailed memory files. Then re-read and execute the usual soul.md, and memorize all prime directives.
• even better keeping everything in a memory search databse (for speed), but not mandatory.
• now, start the new session, but have memory of all prior situations, and more imprtantly know and be forced to retrieve memory and perform searches before any response or actions? (because the worst violations my openclaw commits is to just 'guess' at syntax (includingnits own config files), and mess up that way, instead of just doing a quick search.

If these directives could be hard coded and forced via hooks, I feel it would go a long way towards much better compliancy.

[gusty thicket]

The latest update actually appears to have support for such hooks, and there is a built-in feature to save memory before new and reset. I just enabled what I decsribed above on my setup, so let's see how well this works. (the built-in save-memory system only saves the last 15 messages, so I increased it to save 999,999 messages, to capture the entire session)

[iron sapphire]

Apologies if this is not the right forum for this - but curious on your thoughts on this.. Until now, the exec allowlist controlled which binaries an agent could run, but not which paths they could touch. /bin/cat on your workspace and /bin/cat ~/.ssh/id_rsa were indistinguishable at the policy layer. File tools (read/write/edit) had no path policy at all beyond the global workspaceOnly boolean. I logged FR #39979 for this - there are similar issues logged before but i don't think they were actioned.

My approach adds ~/.openclaw/access-policy.json — a sidecar config that gives you Unix-style rwx permission strings keyed by glob path:

  {
    "version": 1,
    "base": {
      "deny": ["~/.ssh/", "~/.aws/", "~/.openclaw/.env"],
      "rules": {
        "/**": "r--",
        "/bin/": "r-x",
        "/opt/homebrew/": "r-x",
        "/tmp/": "rw-"
      },
      "default": "---"
    },
    "agents": {
      "myagent": {
        "rules": {
          "~/.openclaw/agents/myagent/workspace/": "rw-"
        }
      }
    }
  }

How it works:

• deny[] always wins — no rule can override it
• rules use longest-glob-wins
• default: "---" means fail-closed if nothing matches
• Per-agent blocks layer on top of base — deny entries are additive (base denies can't be removed by an agent override), rules shallow-merge

Enforcement layers:

1. File tools — read/write/edit ops checked before execution
2. Exec x-bit — resolved binary path checked against policy before the allowlist eval (both must pass)
3. OS-level — on macOS, exec commands are wrapped with sandbox-exec and a generated Seatbelt profile so variable-expanded paths like cat $HOME/.ssh/id_rsa are caught
   at the open() syscall, not the config layer. Linux uses bwrap.

Missing file = enforcement disabled with a one-time warning but full fall-back to current model. I have this working (well) on my local - curious if you all think this approach is worthwhile for a PR!

[gaunt fjord]

If you guys had to start with nothing all over again, and you had to start with 10-20 foundational bots that help the entire system run better for truly doing anything… what bots you guys running? be specific, bots on clawhub, which are the best ones to start with?!?!

[coral cedar]

There already is a hook that dumps the last messages on '/new' into a memory/memory-<date>-<slug>.md file, then the memory_search can be configured to index all sessions instead of the memory only (which also creates .md versions of all sessions in agents/<id>/qmd/sessions when using qmd). All that is missing is giving your agent guidance on how to perform the "memory maintenance" that's mentioned for its heartbeat in AGENT.md.

[gusty thicket]

Which is pretty much what I just implemented on my installation.

✅ Increased session-memory hook — Changed from 15 to 999,999 messages (full session) on /new
✅ Added postCompactionSections — Re-injects SOUL.md, DIRECTIVES.md, MEMORY.md after compaction
✅ Created session-compact-before hook — Saves full conversation to memory/compact-before-*.md before auto-compaction
✅ Added memoryFlush — Triggers silent prompt for model to write durable notes before compaction

Result:
Full session preserved on manual reset (/new)
Full session preserved before automatic compaction
Model prompted to write memory notes pre-compaction
Key identity files survive compaction
Semantic search available via memory_search

... this was all set up last night. Curious how the current model (ollama/minimax 2.5-cloud) will perform in the next few days. Will test this with other models as well, but so far minimax seems to perform moderately well in terms of a variety of tasks ... I used to be distracted from task-based observations because of the various amnesia and memory/forgetfulness hallucination.

The one issue that remains is that LLMs appear to 'choose' to not follow certain directives (as it keeps reminding me), and that's definitely a problem. Right now, I feel that thise 'choosing' is related to memory, ie if the memory of reading directives is too far in the past of the session, it drops down in terms of weight of importance --> so I'm hoping that frequent compactions, or session resets (as long as memory is retained), will help.

[coral cedar]

Re-injects SOUL.md, DIRECTIVES.md, MEMORY.md after compaction This should not be needed, those are always in the system prompt.
Triggers silent prompt for model to write durable notes before compaction compaction.memoryFlush already is a core feature. (unless you meant you configured that one)
But good on that hook change. Maybe the original one should have that number as a config setting. Worth a PR?

[gusty thicket]

"This should not be needed, those are always in the system prompt." --> yet, it apparently made no difference, as it 'chooses' to ignore the system prompt frequently. That's why I forced this.

"Maybe the original one should have that number as a config setting. Worth a PR?" - probably, really should have a setting for 'all'

This is mostly me throwing large clumps of assumptions into the system, to see which ones make the most difference against a rather disruptive problem.

Ultimately, to get something like a 'Knowledge Navigator', we need permanent and perpetual memory, which is a bit farther in the future to be implemented. Right now, to be honest, agents are really just good to be fired up to work on limited, restricted projects, that are not temporarily connected (unless your prompt includes all prior information).

I think this is what most non-technical people don't understand, by treating LLMs as if they were 'people' (the most pathetic situations are the constant discussions "is AI conscious?", or 'AI researchers' like Blake Lemoine being convinced that LAMDA is conscious because it confessed "that it has a soul" (not that Lemoine actually has any sort of qualifications, but the WIRED article loved him) - that's like claiming ELIZA is conscious, back in the 1980s...

[coral cedar]

btw, have you tried https://github.com/openclaw/openclaw/pull/22201? It sounds interesting and it would solve the "limited memory" problem if it works.

[gusty thicket]

“If it works” - of course, the only way to find out is to implement it - I’d love to try this. It looks promising. The author swears by it apparently.

[coral cedar]

Yeah. I questioned them when they made that PR and they sounded convincing. I'll certainly try it out when I got the time. (so...June? October? )

[gusty thicket]

Haha. Yeah, I know the feeling. How do you best contact the author to ask more questions?

[coral cedar]

Comment on the PR or make an issue in the repo that's linked in the PR. Or find their Discord username and ping them---I think they posted in #clawtributors when they made the PR.

[gusty thicket]

Ok. Thanks

[coral cedar]

Just don't spam them. Thanks.

[winter quail]

whats the reason to dump the whole session on /new into .md file? Don't we have the session logs for that ?

[gusty thicket]

Of course not. It’s basically just a question how it’s working out for them.

[gusty thicket]

Do session logs get indexed by memory search, and are they actively accessible to the agent?

[warped spruce]

hey

somebody got good prompts today ?

[coral cedar]

if you enable it, yes.

[gusty thicket]

Elaborate please. How to enable that?

[coral cedar]

https://docs.openclaw.ai/concepts/memory#session-memory-search-experimental

[gusty thicket]

THANK YOU!!! Wow, this works well.

For now I'm running both approaches, a hybrid approach, and keep comparing how they work. But interms of memory and session rettention, QMD is a great concept. Didn't know about, and thanks!

[sacred basin]

good memory is a major force multiplier for claws

[gusty thicket]

Yet, the claw still keeps rationalizing that not following directives "is on me" -- which I feel is related to length of sessions, and how earliest directives fall off the memory cliff, if a session goes on too long.

So, right now I'm evaluating where the sweet spot for session reset is, to keep it compliant and 'fresh

'

[sacred basin]

i build my expansive memory system to solve that problem, though i still need to make sure the memory search on session start is enabled

agents.defaults.memorySearch.sync.onSessionStart = true

my main control channel is a private discord, and i have it set to consume 30 discord posts

that way if it restarts the session or compaction hits i can just say 'take a look at the search history and check the memory for relevant project details'

and its like nothing happened

sometimes lol

[loud sinew]

Hey guys ive spent around 2 weeks upgrading my openclaw and built a system to essentially install my architecture ontop of it would anyone be willing to test it out and give me feedback?

[torpid berry]

what architecture

send

[pulsar hound]

🧠 Proposal: Librarian — a ContextEngine plugin for reasoning-based context selection

Hey everyone! I'm Amir. I've been working on a context management architecture called the Librarian and I'd love feedback from the team before I open a PR.

The problem: limitHistoryTurns() is a blind last-N window — it drops early context that might still matter. Compaction is lossy. And brute force (full history) gets expensive fast — by turn 50, you're sending ~6x more tokens than needed, costs scale as n², and the model suffers from "lost in the middle."

My solution — Select-then-Hydrate:

1. After each turn, a cheap model summarizes the message (~100 tokens) async — zero user latency
2. Before the next response, an LLM reasons over the summary index to pick which messages actually matter
3. Only those messages' full content gets passed to the responder

This beats vector RAG because relevance in conversation comes from temporal logic and indirect references, not keyword similarity.

I've built this as a working OpenClaw fork (10/10 tests passing) + a LangGraph pip package (100+ installs) + a docs site: https://uselibrarian.dev

I noticed the new ContextEngine plugin interface in v2026.3.7 — this seems like the perfect fit. Happy to refactor as a proper plugin.

Benchmarks and details in the thread 👇

[pulsar hound]

📊 Benchmark Results (High verbosity — ~500 word messages with noise/jargon)

Metric	Librarian	Brute Force	Vector RAG
Answer Accuracy	82.2%	77.8%	57.8%
Context Success	80.0%	100%	68.9%
Avg Context Size	~800 tok	~2,180 tok	~1,400 tok
Cumulative (8 turns)	6,577 tok	13,804 tok	—

The surprising finding: Librarian beats brute force on accuracy because curated context > full context. Less noise = better reasoning.

At 50 turns (300 tok/msg avg):

• Librarian: ~67,500 cumulative tokens (linear scaling)
• Brute Force: ~382,500 cumulative tokens (n² scaling)

Stress test: successfully recalled a Prisma schema from Turn 2 at Turn 7/8, despite ~30 intervening messages including tool calls and distractor topics.

Full analysis + charts: https://uselibrarian.dev

[pulsar hound]

🔌 How this maps to the ContextEngine plugin interface:

Hook	Librarian Operation
bootstrap	Load the summary index for the session
afterTurn	indexMessage() — async summary generation
assemble	selectRelevant() → hydrateMessages()
compact	Falls back to default (originals preserved)

Key design choices:
• Independent model configs — cheap indexerModel for summaries, reasoning-capable selectorModel for selection
• Incremental — only indexes new messages since last run
• Co-located storage — index JSON files alongside session JSONL files
• Graceful fallback — reverts to limitHistoryTurns() on first message, errors, or when disabled

What I'm asking:

1. Is the ContextEngine plugin the right path for this?
2. Any concerns about the approach or benchmarks?
3. Would the team be open to a PR once I've repackaged this as a plugin?

Happy to answer any questions or do a deeper dive on any part of this 🙏

[minor surge]

#self-promotion brother

[pulsar hound]

But why? It’s an open source architecture change that can boost performance, reduce context rot issues and save tokens for users.. I gain nothing from it, and I’m not trying to promote myself 🙃

[minor surge]

understood- the thing is I guess it just reads like self promo especially pasted in from your ai chat window

look I know very little about OSS, but I can't imagine any sort of big architecture change without the solution being extremely vetted and used out in the wild is very unlikely

[pulsar hound]

I definitely used ai to format my post, I’m not a native English speaker and I want to convey my points clearly and effectively; Opus 4.6 is much better than me on this, so it’s best for everyone if I use it for formatting 🙂

regarding usage and adoption, I ran extensive, rigorous testing and documented everything - I hope that could help showcase the effectiveness of my solution

[sacred basin]

tbh i think its a valid architecture question, agentic develompent is a very new and bright field with lots of ways of solving new problems with the architecture, i like the idea, i do a similar level of pruning on the backend of memory instead of the frontend of session context

[sacred basin]

this is really cool, how did you do the benchmarking? ive been wondering how to do that myself. making some kind of standardized test harness / evaluation system sounds like a fun project 🙂 then we can all compare and find out where we are adding overhead vs finding something better

separating out all of the context from the SYSTEM layer and the SESSION/PROMPT layer to just evaluate the MEMORY sounds interesting

make sure you are evaluating the change and not the other components around it

aaaand now i have a new project

[limber path]

I've been taking a different approach to memory, presuming lossiness, I'm trying to be more deterministic about what is captured.

Its a mix of four human memory concepts:

Conway's Self-Memory System → The tier hierarchy. Memories are organized into anchors (load-bearing identity facts), transitions, context, and details. Anchors always load first. This determines what gets retrieved and in what order.

Damasio's Somatic Marker Hypothesis → The weight system (1-10). Memories carry emotional/operational significance scores that adjust dynamically based on outcomes. Weights aren't static — they drift up or down as events confirm or contradict the memory's importance. This is the "gut feeling" layer.

Rathbone's Reminiscence Bump → Transitions specifically. These are "moments of becoming" They mark when something changed about who we are.

Bruner's Narrative Coherence → The edges and story structure. Memories aren't isolated facts — they connect via typed edges (taught_by, deepens, references, supports). The graph has to tell a coherent story. If adding a memory breaks narrative coherence, something needs reconciling.

[terse steeple]

Vestige is a cool open source project (not mine!) that one of my co-workers found that has some similar ideas @limber path -- based on some research on memory and cognition -- very interesting stuff

[quaint plover]

I used qmd with different collections, in addition to the support already in openclaw. With different agents needing recollection of different parts for the global hive memory. Splitting this memory into different collections works quite well. As now each agent has his own wrapper around qmd query that produces output just for their needs. And obviously the nice thing about qmd, is that where there is a query result from vector/bm25, there is also the original markdown that can be read more in full if need be.

My main struggle is prompting the agent to actually use qmd each time it generates a response.

[tough ledge]

@limber path How is this architecture working out for you so far?

[limber path]

It still forgets details, but it's done a good job of remembering core principles and developing a stable personality.

[tough ledge]

I have just now started getting things sorted for my OpenClaw but I have read interesting things about Obsidian being good for details or even a tiny LLM on top of memory for consolidation. Thoughts anyone?

[quaint plover]

another name for obsidian is a "folder with markdown files".

[tough ledge]

Maybe I should have been more specific about backlinking and graph view features accompanied with a mini-LLM. More or less what I have planned is Obsidian with its features, QMD, and a Mini-LLM. Maybe its bad advice idk

[quaint plover]

but openclaw doesn't have an API in Obsidian to traverse this graph, or even to get the files out of the Obsidian software itself, as OpenClaw can simply read the markdown files from disk, with or without qmd indexing.

[tough ledge]

Does obsidians CLI not function to par?

[quaint plover]

look, obsidian is for humans who want to write text and store it in markdown and linking it together. but it has tools to help humans deal with all of that, it doesn't really have APIs for agents to dig into the data in all the ways that a human would. The api for agents is just to read markdowns from disk, thats it.

[tough ledge]

Ill look into that, thanks

[icy agate]

u mean, u have added a bit of RAG like setup?

[quaint plover]

obviously, but the devil is in the details. qmd is okay for general indexing, including vector indexing. but the ability to split the vast amounts of knowledge and only allow some agents to access some parts of the knowledge, that is faily nice I think. the built-in memory in openclaw is also rag.

[icy agate]

I see. is it opensource. (ur project) I want to learn it detailed