MetaClaw: Self-Evolving Agent Framework That Learns from Every Conversation

MetaClaw is a new open-source agent framework from AIMING Lab at UNC-Chapel Hill that enables continuous learning through real-world conversations — no GPU clusters required. Every live conversation becomes a learning signal, allowing the agent to continuously improve through deployment rather than offline training.

The framework operates through three core mechanisms: Skill Injection retrieves relevant skill instructions and injects them into the agent's system prompt for immediate behavior improvement. Skill Evolution analyzes each session and distills new skills automatically, with RL-enabled judge models extracting skills from failed episodes. A Meta-Learning Scheduler defers weight updates to idle windows so the agent is never interrupted during active use.

MetaClaw supports three operating modes: Skills Only (proxy-based, no GPU needed), RL Mode (continuous fine-tuning), and MadMax Mode (skills + RL with intelligent scheduling). It's compatible with multiple coding agents including OpenClaw, CoPaw, IronClaw, and NemoClaw.

The technical report was released on March 18, 2026, trending on HuggingFace Daily Papers with 75 upvotes. The project has ~2,000 GitHub stars under MIT license.

GitHub: https://github.com/aiming-lab/MetaClaw