Index / MetaClaw · updated Oct 24, 2026
MetaClaw
aiming-lab/MetaClaw · healthy · rank 22 of 50 by stars
A research-grade OpenClaw companion that meta-learns and evolves skills from every conversation, no GPU cluster required. It layers cross-session memory and scheduled RL training on top of the OpenClaw agent experience.
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 — 11 months old (from the public model catalogue)
Pin last edited 5 months ago
Evidence
Decision
Why choose MetaClaw over OpenClaw?
Why choose this
- Automatic skill evolution from conversations
- Cross-session user/project memory injection
- Optional RL training mode without GPU cluster
Tradeoffs
- Depends on OpenClaw rather than replacing it
- Younger project with research-grade stability
- Less focus on sandboxing and hardening
Best fit
- Researchers experimenting with agent self-improvement
- OpenClaw users wanting persistent cross-session memory
- Teams benchmarking RL-based skill evolution
Avoid if
- You need a hardened, sandboxed production agent
- You want a mature plugin ecosystem and docs
- You avoid research-stage APIs that change frequently
Strong evidence from README, releases, and detailed commit history that this is an active OpenClaw-adjacent project. Community sentiment is thin and partly noisy, so sentiment and risk scores carry moderate uncertainty.
AI layer reviewed Oct 5, 2026 · how this is written
Star activity
3,465 stars today
Overview
MetaClaw is a Python-based meta-learning layer for OpenClaw-style agents: you talk to your agent and it learns and evolves skills from every conversation, with no GPU cluster required. It offers three modes — skills-only, RL (via Tinker), and a default auto mode combining skill evolution with scheduled RL training — plus a 'Contexture' memory layer that persists cross-session user and project facts and injects them into prompts.
Architecturally it ships as a CLI (metaclaw setup / metaclaw start), an async API server and proxy, and an OpenClaw memory sidecar plugin with endpoints like /buffer_turn and /flush_session for incremental per-turn memory ingestion (added in v0.4.1). Recent commits show heavy investment in a benchmark suite (metaclaw-bench), Windows cross-platform fixes, and configurable context windows, indicating active, engineering-driven development.
Unlike OpenClaw, which is the canonical personal-assistant runtime, MetaClaw positions itself as a learning/evolution companion rather than a replacement — its value is continual self-improvement and memory, while sandboxing, hardening, and ecosystem breadth remain OpenClaw's strengths.