aiming-lab/MetaClaw

Implements meta-learning through conversation interception—MetaClaw proxies OpenAI-compatible APIs (OpenClaw, NanoClaw, CoPaw, IronClaw) and injects learned skills at each turn, auto-summarizing sessions without requiring GPU infrastructure. Supports three modes (skills-only, RL-only, or MadMax with scheduled training), asynchronous reward modeling and LoRA updates via Tinker/MinT backends, and cross-session memory that automatically retrieves and injects user preferences and project history. Default MadMax scheduler defers weight updates to idle windows, sleep hours, or calendar meetings to avoid interrupting live conversations.

805 stars. Actively maintained with 171 commits in the last 30 days.