MetaClaw: Just Talk — An Agent That Meta-Learns and Evolves in the Wild (2026)

Free

Continual meta-learning framework that jointly evolves a base LLM policy and a reusable skill library — skill-driven fast adaptation from failure trajectories + opportunistic gradient updates during idle periods; 21.4% → 40.6% accuracy on benchmarks (134 HF likes)

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About MetaClaw: Just Talk — An Agent That Meta-Learns and Evolves in the Wild (2026)

MetaClaw is a continual meta-learning framework for LLM agents that jointly evolves a base LLM policy and a library of reusable behavioral skills. It uses two complementary mechanisms: skill-driven fast adaptation analyzes failure trajectories via an LLM evolver to synthesize new skills, enabling immediate improvement with zero downtime, and opportunistic policy optimization performs gradient-based updates via cloud LoRA fine-tuning and reinforcement learning with a process reward model, triggered during user-inactive windows by the Opportunistic Meta-Learning Scheduler (OMLS). A versioning mechanism prevents data contamination, and a proxy-based architecture scales to production-size LLMs without local GPUs. Experiments on MetaClaw-Bench and AutoResearchClaw show skill-driven adaptation improves accuracy by up to 32% relative, advancing Kimi-K2.5 accuracy from 21.4% to 40.6% and increasing composite robustness by 18.3%.

Key Features

Jointly evolves a base LLM policy and a library of reusable behavioral skills

Skill-driven fast adaptation: synthesizes new skills from failure trajectories via an LLM evolver

Opportunistic policy optimization: gradient-based updates via cloud LoRA fine-tuning and RL-PRM during idle periods

Opportunistic Meta-Learning Scheduler (OMLS) monitors system inactivity and calendar data

Versioning mechanism separates support and query data to prevent contamination

Proxy-based architecture scales to production-size LLMs without requiring local GPUs

Achieves up to 32% relative accuracy improvement and 18.3% increase in composite robustness

Pros & Cons

Pros

Zero-downtime adaptation via skill synthesis from failure trajectories
Significant accuracy improvement demonstrated (21.4% to 40.6% on Kimi-K2.5)
Increases composite robustness by 18.3%
Scalable to production-size LLMs without local GPU requirements
Combines fast skill adaptation with deeper policy optimization for mutual reinforcement

Best For

Continuous adaptation of LLM agents on platforms handling diverse workloads across multiple channels (e.g., OpenClaw with 20+ channels)Deploying LLM agents that need to evolve without downtime as user needs changeMeta-learning and skill library management for production-grade LLM agents