MetaClaw is an asynchronous agent that learns from conversations, supports multiple open-source claws, and can persist memory across sessions. It provides a one-click deployment flow, RL options, and configurable backends including Tinker, MinT, and Weaver.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
MetaClaw is an agent framework that meta-learns and evolves from live conversations. It acts as a proxy in front of a personal agent (OpenClaw, CoPaw, IronClaw, PicoClaw, ZeroClaw, NanoClaw, NemoClaw, or any OpenAI-compatible client) to inject skills at each turn and enable continual improvement without requiring a GPU cluster.
How it works
- Serves as a proxy in front of a personal agent, injecting skills at each turn and aggregating experience for meta-learning.
- Memory layers support long-term persistence of facts, preferences, and project history across sessions.
- Asynchronous design: serving, reward modeling, and training run decoupled; agent responses can continue during optimization.
- RL path supports different backends (Tinker, MinT, Weaver) chosen via configuration.
- Optional OpenClaw, CoPaw, Hermes, and other claws can be auto-configured at start; memory and prompts are injected per turn.
Getting started
- Quick Start section provides installation commands for OpenClaw and Python options.
- One-click deployment flow: run metaclaw setup, then metaclaw start. Examples:
metaclaw setup
metaclaw start
- OpenClaw installation example (v0.4.0):
curl -LO https://github.com/aiming-lab/MetaClaw/releases/download/v0.4.0/metaclaw-plugin.zip
unzip metaclaw-plugin.zip -d ~/.openclaw/extensions
openclaw plugins enable metaclaw-openclaw && openclaw gateway restart
- Python package installation examples (for different modes):
pip install -e . # skills_only mode (lightweight)
pip install -e ".[rl]" # + RL training support (torch, transformers, tinker)
pip install -e ".[evolve]" # + skill evolution via OpenAI-compatible LLM
pip install -e ".[scheduler]" # + Google Calendar integration for scheduler
pip install -e ".[rl,evolve,scheduler]" # recommended for full RL + scheduler setup
- WeChat integration steps are provided if enabled via metaclaw config and start.
Getting started (continued)
- Configuration wizard asks to choose agent, auth method, LLM provider, and model, with options for rl.backend (tinker/mint/weaver) and other RL settings.
- Start command:
metaclaw start
- Quick start for multiple claws and memory features is complemented by the Multi-Claw Support section detailing how each claw is wired at startup.
Recent releases
- v0.4.1 (2026-04-11): Incremental memory ingestion – memory layer persists turns every N turns (default 5) instead of only at session end.
- v0.4.0 (2026-03-25): Contexture layer – cross-session memory persistence for users/projects; prompts injected with relevant facts and preferences.
- v0.3.3 (2026-03-23): One-click OpenClaw plugin – MetaClaw ships as a native OpenClaw extension.
- v0.3.2 (2026-03-23): Hermes Agent support; configurable context window; memory and agent integration improvements.
- v0.3.1 (2026-03-14): MinT backend support; RL training with Tinker/MinT; backend abstraction.
- v0.3 (2026-03-13): Continual meta-learning support; delayed RL updates to idle times.
- v0.2 (2026-03-11): One-click deployment via metaclaw CLI; skills enabled by default; RL opt-in.
- v0.0 (2026-03-09): Initial release of MetaClaw.
Traction
- Stars: 3514
- Forks: 452
- Open issues: 17
Configuration
Configuration lives at ~/.metaclaw/config.yaml, created by metaclaw setup. CLI commands include setup, start, stop, status, config show, and various rl.backend/model-related settings. The README provides explicit command examples for configuring rl.backend (tinker, mint, weaver) and associated API keys and models.