Personal Productivity

How aiming-lab Built MetaClaw So OpenClaw Agents Learn from Chats

Built by aiming-lab (research team) · · Source: GitHub/arxiv

Quick answer

aiming-lab built MetaClaw, an open source proxy that sits between an agent like OpenClaw and its model. It injects relevant skills at each turn, writes new skills after each session, and can fine-tune the model from scored conversations during idle hours.

Result: About 3.5k GitHub stars and 455 forks. The README says the technical report ranked No. 1 on HuggingFace Daily Papers. It gives no benchmark scores.

What they built

MetaClaw is an open source, MIT-licensed project from the aiming-lab organization on GitHub. It is a proxy that sits between your personal agent and the model behind it. According to the README, the proxy intercepts the agent's requests, injects relevant skills at each turn, and learns from the experience that builds up over time. The team says it needs no GPU cluster and works with any OpenAI-compatible LLM API.

OpenClaw is the primary supported agent. The README also lists CoPaw, IronClaw, PicoClaw, ZeroClaw, NanoClaw, NemoClaw and Hermes Agent. The project comes with a paper (arXiv 2603.17187) by Peng Xia and twelve co-authors.

How it works

  • Skills: Skills are short Markdown SKILL.md files stored in ~/.metaclaw/skills/. The README says new ones are summarized automatically after each session, and the proxy adds the relevant ones to each turn.
  • Three modes: skills_only runs the proxy and skill injection with no GPU. rl adds continuous reinforcement learning. auto, the default, adds a scheduler on top of both.
  • Training: In RL mode each turn becomes a training sample. A judge LLM scores responses in the background, and a LoRA backend (Tinker, MinT or Weaver) fine-tunes the model and swaps in the new weights. A separate evolver LLM pulls new skills out of failed episodes.
  • Scheduler: In auto mode, weight updates wait for sleep hours, keyboard inactivity or Google Calendar meetings, so the agent is not interrupted while you are using it.
  • Memory: Version 0.4.0 added a long-term memory layer that stores things like preferences and project state and injects relevant memories each turn.

Setup is a pip install from the repo, then metaclaw setup and metaclaw start. Settings live in ~/.metaclaw/config.yaml. For OpenClaw, the README also offers a one-click plugin from the v0.4.0 release.

The result

The README does not report performance numbers. It mentions MetaClaw-Bench, a benchmark built on multi-day interaction histories, but gives no scores. It does say the technical report ranked No. 1 on HuggingFace Daily Papers. The repo shows about 3.5k stars and 455 forks, and the news section lists releases from the first version on 03/09/2026 through v0.4.1 on 04/11/2026.

How to build something similar

  • Start with the skills part only. Have your agent write a short lesson note after each session and load the relevant notes into the next one. This needs no training setup.
  • Put a proxy between the agent and the model. That gives you one place to log every turn and add context without changing the agent itself.
  • Read the generated skill files regularly and delete the ones that are wrong. Bad lessons get repeated in every later session.
  • Only add fine-tuning once you have plenty of logged turns and a way to score them, and run it when you are not using the agent.

The original post

aiming-lab · View on GitHub/arxiv

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases