#HindClaw — Production Memory Infrastructure for AI Agents

1 messages · Page 1 of 1 (latest)

alpine berry
#

https://github.com/mrkhachaturov/hindclaw
https://hindclaw.pro — production memory infrastructure built on Hindsight

On March 17 I asked Perplexity how to improve memory for my AI agents. It pointed me to Hindsight. Five days later I shipped this:

Three packages, one platform:

hindclaw-extension (PyPI) — server-side Hindsight extensions. JWT auth, user/group permissions, tag-based recall filtering, retain strategy enrichment. All access control lives on the server, not the client.

terraform-provider-hindclaw (Terraform Registry) — manage users, groups, banks, permissions, directives, mental models, and entity labels as code. terraform apply and it's live.

hindclaw-openclaw (npm) — OpenClaw gateway plugin. Thin adapter — signs JWTs per message, auto-starts embed daemon with extensions loaded. One install, zero manual setup.

What it does:

  • Per-agent memory banks with custom missions, entity labels, dispositions, and directives
  • Server-side permission cascade: global defaults → group → bank group → bank user override
  • Tag-based recall filtering and retain strategy enrichment via accept_with()
  • Multi-bank recall — one agent reads from multiple banks in parallel
  • Mental models and session start context
  • Multilingual entity labels with controlled vocabulary

Running 11 agents on Telegram. Fully self-hosted, fully open source.

Built on https://hindsight.vectorize.io — the highest-scoring agent memory on LongMemEval. They build the engine, HindClaw adds what you need to run it in production with multiple users and agents.

soft drift
#

ive been grinding the memory grind for a while and ive got to say hats off

#

theres a few things i was going for in my next revisions that you've hit

#

so much exciting potential here

dark folio
#

its a saas backend wrapper

#

why pay for memory when ClawMem does all this (and more) and is oss?

alpine berry
#

HindClaw — Production Memory Infrastructure for AI Agents

queen spade
dark folio
# queen spade isnt this their saas github if you want to self host ? https://github.com/vector...

ur actually right, they do offer fully self-hosted with MIT license. i wasn't aware of this so thanks for pointing that out. its actually quite sophisticated. the one flaw is that it still relies on agents to proactively make use of tools (whatever way you wrap them) to do the retrieval. this is the bottleneck for all proactive memory tooling. models rarely call manual memory tools and fallback to native tooling even when explicitly instructed to use the memory tools.

#

ideally a wrapper for this uses hooks (so u dont need to rely on the LLM for retrieval) and for OpenClaw this means ContextEngine plugin with hooks mapped to the event-based stuff that OpenClaw exposes.

#

u can even force the graph traversals with hooks, though how deep u want to go on a retrieval with hooks would depend on several factors

#

this is why ClawMem uses 3 tiers of depth with tweakable timeouts, etc

#

quite a few goodies in hindsight to wrap or extract so thanks for the heads up

alpine berry
#

@LarpsAI That's exactly what HindClaw does — it uses OpenClaw's hook system, not LLM tool calls. The plugin hooks into before_prompt_build for recall and agent_end for retain. No memory tools exposed to the model, no reliance on the LLM deciding to call them.

The recall query is composed from conversation context automatically and injected into the system prompt before the LLM even sees it. Retain fires after every turn. The model never knows memory exists — it just gets richer context.

On self-hosting: Hindsight is fully open source (MIT), runs locally via hindsight-embed with an embedded PostgreSQL (pg0) — no external database needed. HindClaw's plugin auto-starts the daemon, auto-installs the server extensions, auto-configures JWT auth. One plugin install, openclaw gateway, and it's running.

For Docker deployments, Hindsight has a standard hindsight-api image that runs against any PostgreSQL. The hindclaw-extension installs as a pip package into the same environment. Same extensions, same access control — just a different deployment model.

The full stack: embed for dev/single-user, Docker for production, Terraform for managing the infrastructure layer on top.

dark folio
#

hook-only is good enough for OpenClaw i suspect

alpine berry
dark folio
#

agent_end too.. saves complexity of needing an LLM involved but probably tends to get a bit noisy if just doing convo dumps?

#

hindsight approach def good for multi-user

alpine berry
# dark folio agent_end too.. saves complexity of needing an LLM involved but probably tends t...

Good point — but Hindsight's retain isn't a raw dump. The conversation goes through LLM-powered fact extraction with a retain_mission that tells it exactly what to keep and what to ignore. So even though agent_end sends the full transcript, the engine extracts only relevant facts based on the bank's mission.

For example, the health agent's mission says "Extract sleep, workout, nutrition data. Ignore: work topics, financial data." Same conversation, different agents extract completely different facts from it.

Entity labels add another filter — the LLM classifies each fact into a controlled vocabulary (person, company, scope tier) and only stores what matches the schema. Noise gets dropped at extraction time, not at ingestion time.

The retainRoles config also helps — you can retain only ["user", "assistant"] and skip tool calls, system messages, etc. And retainEveryNTurns lets you batch (e.g., retain every 3rd turn with overlap) instead of every single turn.

dark folio
#

thank your LLM for the reply bro xD

alpine berry
#

english is not my native language bro

dark folio
#

all g, try it with that skill tho