From Memory to Skills: Evidence-Grounded Co-Evolution Governance for Long-Horizon LLM Agents
Computation and Language
Summary
The authors propose a new system called MSCE that helps AI agents remember and use what they've learned more effectively. Instead of just storing past information as plain memory, MSCE turns these experiences into useful skills the agent can call upon later, with built-in checks to know when and how to use them. They also introduce a method to better learn from feedback by reflecting on actions step-by-step. Their tests show MSCE works better than other recent methods, even when facing different tasks over time.
Authors
Bo Tang, Yang Zhang, Guomian Zhuang, Wenqiang Wei, Gaoyang Zheng, Lindong Xie, Yanchao Tan, Feiyu Xiong, Qingyu Yang, Edward Chung, Zhiyu li
Abstract
Existing memory systems for long-horizon LLM agents often retrieve prior traces as passive context rather than converting them into executable capabilities. In this paper, we propose MSCE, a training-free Memory--Skill Co-Evolution framework that organizes agent experience into grounded step traces, reusable procedural policies, and declarative environmental cognition. MSCE crystallizes evidence-backed L2 policies with positive estimated gain into callable skills that retain evidence links, applicability boundaries, decision guidance, verification rules, and reliability estimates. It further introduces reflection-weighted value backfilling, which propagates sparse terminal feedback through dense local self-reflections to produce evidence-calibrated trace values for governing memory and skill evolution. Experiments on EvoAgentBench and LoCoMo demonstrate that MSCE significantly outperforms state-of-the-art skill-augmented and memory-driven agent baselines, exhibiting strong cross-domain transferability and lifelong-evolution capabilities.