Memento-Skills Framework: Agents Rewrite Their Own Skills via Read-Execute-Reflect-Write Loop — 80% Task Success vs 50% RAG Baseline, No Model Retraining Required

VentureBeat·medium signal

Memento-Skills treats skills as first-class mutable units of capability in an evolving external memory. The core loop: Read (retrieve relevant skill), Execute (run it), Reflect (evaluate outcome), Write (improve and store back). When a task fails, the system identifies the weak skill, improves it, and writes the improvement back — all without updating LLM parameters. On General AI Assistants benchmark: 26.2% relative accuracy improvement; Humanity's Last Exam: 116.2% improvement. 80% task success vs 50% for standard RAG retrieval. Ships with CLI, GUI, and Feishu integration plus local sandbox execution.