Official code for the paper Rethinking Continual Experience Internalization for Self-Evolving LLM Agents.
The pipeline turns an agent's interaction trajectories into a compact, principle-level experience pool and internalizes it into the model parameters, so the agent and its experience co-evolve across iterations.
experience_extraction/— agent rollout + experience summarization: a local model solves tasks (web reasoning / math), and the scored trajectories are distilled into a reusable experience pool. See its README for setup and usage.evaluation/— web-agent evaluation on WebWalkerQA, GAIA-Text-103, and BrowseComp-ZH, with and without inference-time experience injection. Built on the Tongyi DeepResearch inference framework. See its README.
More components (experience injection and context-distillation training) will be released here.