Large Language Model (LLM) agents achieve impressive results on complex tasks, yet they usually operate in isolated sessions and cannot continuously learn from past interactions or adapt to user‑specific environments. Existing self‑evolution approaches often inject raw interaction trajectories directly into memory or skill modules, obscuring behavioral patterns and reducing update controllability. As foundational LLM reasoning improves, the bottleneck shifts from accumulating generic knowledge to managing situational experience effectively. To address this, we introduce Darwin-Claw, an environment‑adaptive self‑evolving framework that systematically distills episodic experiences to drive comprehensive evolution of operational skills and durable memory. The framework is grounded in an iterative co‑adaptation between an agent‑actor and an evolver‑critic, and features a Hierarchical Experience Consolidation System that extracts fine‑grained error attributions and abstracts them into dynamically updated prototype vectors. These structured situational experiences serve as interaction patches for context‑sensitive perception and provide explicit optimization signals for skill refinement. Darwin-Claw can be deployed out‑of‑the‑box with mainstream agent systems without manual curation. Extensive experiments on WildClawBench demonstrate substantial performance gains, establishing a controllable and stable pathway for continual, user‑centric adaptation of autonomous agents in real‑world ecosystems.
Review