Has anyone used qmd?

[sweet kiln]

I saw this project that seems promising for managing the memory files.

Has anyone tried OpenClaw with a plugin or skill to use that? https://github.com/tobi/qmd

[robust vigil]

I'm using, but my Claw....less so. Have to check on it regularly.

[obsidian kindle]

I’ve been using https://clawvault.dev/ in an Obsidian vault. It utilizes qmd and can help manage the setup / intelligent use with a basic / semi simple set up.
Paying for the obsidian sync has allowed me to jot ideas from my phone or windows PC when needed as well as making a nice secure backup of my claw brain. 🧠

[elfin flame]

i integrated QMD in https://github.com/yoloshii/ClawMem but with the option of using gguf versions of SOTA models for embedding and reranking (zembed-1 and zerank-2) with a gpu

QMD native models are good for apple silicon or anything with integrated graphics, but users without that it goes to cpu only which is way too slow to be useful.

[compact crown]

It works great and is good for searching your claw’s memories. I told mine to install it and make it the primary method for searching memories. Once you get that installed and working next tell it to install LCM (lossless-claw) it will give your claw memory over reboots, upgrades, you name it. https://github.com/Martian-Engineering/lossless-claw

[proud crypt]

How can we tell when these memory systems are working? I have qmd setup and lossless-claw, but I don’t have hard proof that they are working. I think they are working. Openclaw says they are setup correctly and being used, but I’d like to have some sort of log or visual representation of these memory systems in use.

[compact crown]

Ask it some questions from memories from days or weeks ago and see if it remembers? Then tell it some fact about you and immediately do a openclaw gateway restart then ask it what the fact was before you ran the restart. If it can tell you then LCM stored it and claw was able to retrieve it.

[elfin flame]

ask the model to analyze if it gets any retrievals and how relevant they are, etc

[proud crypt]

I’ll try both of these. Thanks

[snow peak]

Can someone show the config about the qmd, I have lots of troubles in install it.

[compact crown]

Tell your claw to install it & configure, way easier that way. But you can run npm install -g @tobilu/qmd then add it to your openclaw.json and restart the gateway.

"memory": {
����"backend": "qmd",
"citations": "auto",
"qmd": {
������"command": "qmd",
������"includeDefaultMemory": true,
������"paths": [������],
������"sessions": {
��������"enabled": true,
������ "retentionDays": 30
},
"update": {
�������"interval": "5m",
��������"debounceMs": 15000,
��������"onBoot": true,
��������"embedInterval": "1h"
������},
������"limits": {
��������"maxResults": 6,
��������"maxSnippetChars": 700,
��������"maxInjectedChars": 4000,
��������"timeoutMs": 4000
������},
������"scope": {
��������"default": "deny",
��������"rules": [
����������{
������������"action": "allow",
������������"match": {
��������������"chatType": "direct"
������������}
����������}
��������]
������}

[viscid sage]

QMD is a solid option if local-only is a hard requirement for you.

I went a different route with mr-memory (openclaw plugins install mr-memory) which takes a different approach — cloud-based vector storage. The tradeoff is it's not local, but what you get for that: instant retrieval speed regardless of how much data you have (parallelized), infinite storage so nothing gets summarized away, and you keep the actual nuance of your conversations, not just high-level summaries. Zero config, works from day one.

Different philosophies — local control vs. speed + depth. Depends on what matters more to you.

[snow peak]

I want set another model for embeding . And can I set remote models for rerank and query model. If I can't use the remote models , Can I just use the embeding model only.

[elfin flame]

i have this type of setup with: https://github.com/yoloshii/ClawMem You can use SOTA embeding and reranking models (zembed-1 and zerank-2), or any model you want if u put it behind the endpoint. Or u can use cloud embedding models.

[gray heart]

I'm currently testing a memory-based skill/plugin with multiple LLMs.
The memory is organized in multiple layers with .md files and a local database, because I prefer to keep my files on my own device. The memory layers work between LLMs to pass information, and they update in real time. One advantage is that you can configure the LLMs used in the memory layers for optimization.
If you're interested => https://clawhub.ai/nieto42/openclaw-memoria

[snow peak]

I have read the readme.md, it only contains
CLAWMEM_EMBED_URL
CLAWMEM_EMBED_API_KEY
CLAWMEM_EMBED_MODEL
CLAWMEM_LLM_URL
CLAWMEM_RERANK_URL

but what about
CLAWMEM_ LLM_API_KEY
CLAWMEM_ LLM_MODEL
CLAWMEM_ RERANK_API_KEY
CLAWMEM_ RERANK_MODEL
？ I don't know whether it supports setting a cloud model for LLM amd RERANK.

[elfin flame]

give it to your LLM and it will tell you

only embed model from 4 providers in cloud

if you want to add more just add them