MetaClaw: an agent proxy that enables AI assistants to meta-learn and evolve through real-world conversations

What it solves

MetaClaw addresses the problem of AI agents remaining static after deployment. Instead of relying solely on offline training, it allows an agent to continuously learn and evolve from real-world conversations in the wild, improving its performance and adapting to user preferences over time without requiring the user to manage a GPU cluster.

How it works

MetaClaw acts as a transparent proxy between a personal agent (such as OpenClaw, CoPaw, or NanoClaw) and an LLM API. It intercepts interactions to inject relevant skills and persist long-term memory.

Depending on the mode, it handles learning in different ways:

• Skills Mode: Automatically summarizes conversations into short Markdown instructions (skills) that are retrieved and injected into future prompts.
• RL Mode: Uses a judge LLM (PRM) to score responses asynchronously and performs LoRA fine-tuning via cloud-based backends like Tinker, MinT, or Weaver.
• Auto Mode: Combines skills and RL, using a smart scheduler to defer weight updates to idle windows (sleep, idle time, or calendar meetings) so the agent is never interrupted.

Who it’s for

It is designed for users of personal AI agents who want their assistants to evolve and remember facts, preferences, and project history across sessions without needing local high-end hardware.

Highlights

• One-click deployment: Simple CLI setup that automatically configures supported personal agents.
• Multi-agent support: Compatible with a wide range of agents including OpenClaw, CoPaw, IronClaw, and others via OpenAI-compatible or Anthropic-native endpoints.
• Long-term memory: Persists cross-session facts and preferences to provide consistent context.
• Asynchronous architecture: Decouples serving, reward modeling, and training to ensure zero latency during active use.
• Flexible RL backends: Supports multiple cloud training providers including Tinker, MinT, and Weaver.