Index / MetaClaw · updated Oct 2, 2026
MetaClaw
aiming-lab/MetaClaw · healthy · rank 18 of 43 by stars
A research-grade OpenClaw companion that meta-learns and evolves skills from every conversation, no GPU cluster required. It layers persistent cross-session memory and scheduled RL training on top of the OpenClaw agent loop.
Facts
Repository
Runtime
Posture
Community
Security breakdown
Composite 55 / 100 · how these are scored
higher is safer
higher is safer
higher is safer
higher is safer
higher is riskier
Model access
Read from the repository, not written by a model · 23 files examined
2 providers · custom endpoint
Pinned models
gpt-5.2 released 2025-12-10 — 10 months old (from the public model catalogue)
Pin last edited 5 months ago
Evidence
Decision
Why choose MetaClaw over OpenClaw?
Why choose this
- Built-in skill evolution and RL training loop (Tinker-based)
- Cross-session memory layer with automatic fact retrieval and injection
- One-command setup wizard and async architecture
Tradeoffs
- Depends on OpenClaw as the underlying agent rather than replacing it
- Less mature messaging/channel integrations than core OpenClaw
- Security and sandboxing posture largely undocumented
Best fit
- Researchers experimenting with agent self-improvement and RL
- OpenClaw users wanting persistent cross-session memory
- Teams benchmarking skill-evolution approaches without GPU clusters
Avoid if
- You need a hardened, production-ready sandboxed agent
- You want a standalone assistant rather than an OpenClaw add-on
- You are uncomfortable with early-stage research code and frequent breaking changes
Strong evidence from README, releases, and detailed commit history for architecture and activity claims. Reddit/web matches are mostly noise (unrelated 'meta' hits), so community sentiment is uncertain.
AI layer reviewed Sep 21, 2026 · how this is written
Star activity
3,459 stars today
Overview
MetaClaw is a Python-based meta-learning layer for OpenClaw-style agents, built around the tagline 'just talk to your agent — it learns and evolves.' Its core architecture combines three modes: a Skills mode that extracts reusable skills from conversations, an RL mode that runs scheduled reinforcement training (via Tinker, no GPU cluster required), and an Auto mode that orchestrates both. A 'Contexture' memory layer persists cross-session facts, preferences, and project history, automatically injecting relevant context into prompts, with v0.4.1 adding incremental per-turn memory ingestion to shrink the mid-session memory blackout window.
Unlike a standalone assistant, MetaClaw positions itself as an evolution engine on top of OpenClaw: it ships an OpenClaw memory plugin/sidecar (TypeScript hooks plus a Python memory service), an API proxy, and a full benchmark suite (metaclaw-bench) for measuring agent improvement over multi-day task sequences. Recent commits show active maintenance — Windows cross-platform fixes, config handling improvements, and benchmark data corrections — alongside a technical report on arXiv.
Compared to OpenClaw proper, MetaClaw trades breadth (messaging channels, community plugins, hardened deployment) for depth in self-improvement: skill evolution, scheduled RL, and structured long-term memory. It is best understood as a research-forward companion project rather than a drop-in OpenClaw replacement, and its MIT license and one-command setup (metaclaw setup && metaclaw start) make it easy to evaluate.