#ClawMem - Advanced SOTA Local Memory Plugin

1 messages · Page 1 of 1 (latest)

maiden phoenix
#

Most memory plugins embed and retrieve. ClawMem does that too, but the retrieval is where it really cooks.

Search core is forked from QMD (BM25 + vector + RRF + query expansion + cross-encoder reranking). You can keep the small models that run anywhere, or swap in zembed-1 and zerank-2 on a GPU for dramatically better retrieval. zerank-2 beats Cohere rerank-3.5. Small models stay as automatic fallback.

What no other plugin does (as far as I know):

  • Ask "why did we decide X" and it classifies the intent as WHY, then traverses causal graphs via multi-hop beam search instead of doing keyword search
  • Content-type half-lives. Decisions never decay. Handoffs fade after 30 days. Most plugins treat all memories the same.
  • A-MEM self-evolving memory. Index a new note and it finds related docs, assesses the relationship, writes causal/supporting/contradiction links automatically
  • Contradiction detection. New decision contradicts an old one? Old decision's confidence drops on its own.
  • Co-activation reinforcement. Notes surfaced together repeatedly get a retrieval boost. The vault learns what clusters.

Shares a vault with Claude Code if you use both. Decisions from one runtime show up in the other immediately.

Watcher re-indexes on file changes, embed timer keeps vectors fresh, curator agent sweeps stale memories.

npm install -g clawmem
clawmem setup openclaw

GitHub: https://github.com/yoloshii/ClawMem
npm: https://www.npmjs.com/package/clawmem
Maintainer: yoloshii (LarpsAI)

worldly gorge
#

This sounds extremely efficient. Is anyone else using this and what are your experiences?

#

Especially you, the creator of this, what re your experiences in using this? Does this help vanquish agents suffering from amnesia after a while? Do any other configurations need to be set for models to work best with this?

maiden phoenix
#

Hi, yes it works. The trouble with these agent memory frameworks is you often don't know what is happening under the hood (ie; if its working or if its failing silently) so I will likely build out a monitoring pipeline to show the processes in realtime. Also, people need to realize that all of these filesystem based context engines are only as good as your filesystem. If you have just a few documents here and there, perhaps a couple of memory files, its not going to provide much context that is relevant to the agent's current tasks. So the more you index and add to collections that relates to the usual things you do with your agents, the better it works. Your indexed collections should include not only memory files but context documents, research outputs, notes, learnings, decision points, domain expertise references, etc, so that your agents have a rich field of context for the retrieval to surface at every turn. I normally set the config to index every .md file in a given project. Codebases or code on the other hand, is better indexed in a separate vault or a dedicated repo RAG tool. To answer your question, everything is in the documentation on the repo. The best idea is to give it all to your agent to ingest and then have it guide you on the best way to install and configure everything. If you run into any bugs or issues don't hesitate to create an issue on the Github so it can be addressed.

maiden phoenix
#

For the proactive things the agent should be doing (eg; deeper search after surfaced context is found relevant) redundancy is probably a good idea. Add the instruction block to system prompt, global user CLAUDE/AGENTS.md, skillify it. Agents taking the easy route and falling back to simpler native tools is currently the biggest bottleneck when it comes to trying to get them to be proactive with use of memory and context engine tools (the manual ones anyway). This is why ClawMem has as much automated retrieval through hooks and event-driven reactions as it can, but some things your agent still needs to be proactive about. That is, if you don't want to be doing them yourself manually, which is kind of the whole point. This stubbornness of agents in not being proficient with the proactive type of memory tools probably won't change until the model providers include the use of such tools in their training data.

maiden phoenix
#

There's also a second factor. Models implicitly optimize for response speed. MCP calls add latency and context consumption. The model's cost/benefit calculation favors "answer now" over "search first, answer better." This is a different problem than training data. It's an inference-time optimization bias. Even with memory tools in training data, the model might still skip them if it thinks it can answer "well enough" without the call. tl;dr: none of these memory solutions can work perfectly 'out of the box'. Users should experiment, tweak and refine their harness to integrate with them as best they can.

worldly gorge
#

I've expanded my openclaw setup with memory_search and QMD, and it made a lot of difference. The biggest issue I've been dealing with has been Ollama throwing up context API errors, which kept losing context and memory .... the latest Ollama 0.18 update made A HUGE DIFFERENCE, as they addressed that bug with some automatic houseclaning. Now, it shows the error and then auto-recovers.

So if you're using Ollama, update to 0.18 as if your life depended on it ...

maiden phoenix
#

ClawMem now significantly updated

#

includes useful features extracted from MemPalace and other current approaches

maiden phoenix
#

ClawMem now correctly positions as memory plugin instead of context engine after the changes. Now you can pair it with lossless-claw for the added benefit.

#

Quite a few upgraded features for improved accuracy and clean retrieval with more on the way.

#

ClawMem - Advanced SOTA Local Memory

#

ClawMem - Advanced SOTA Local Memory Plugin

brave sand
#

This would be nice. I build something similar, but used Qdrant, and it works great, but it required a lot of modifying under the hood to keep all of the core Active Memory functionality for Memory search, deaming ect.