MSCE turns agent memories into callable skills for long-horizon LLM systems

rohanpaul_ai · x · 2026-07-26

This paper argues that most agent memory systems fail because they merely store past events and feed them back as context, forcing the model to rediscover the same lesson over and over.
The proposed MSCE framework separates declarative facts, induced procedural policies, and grounded step traces, then promotes recurring evidence-backed patterns into callable skills with explicit triggers, boundaries, expected gain, and reliability checks.
Example: instead of recalling a failed pip install incident, the agent learns a reusable rule for identifying the missing OS package, installing it, and retrying.
The authors also add reflection-weighted value backfilling to propagate sparse terminal feedback through dense local reflections.
Experiments on EvoAgentBench and LoCoMo reportedly show strong gains over state-of-the-art skill-augmented and memory-driven agent baselines, with better cross-domain transfer and lifelong evolution.

More from coding & agent