做过十遍还要从头摸索？Agent 的记忆终于开始长成技能
Merged summary
TL;DR - MSCE is an arXiv framework that turns long-horizon agents’ accumulated experience into evidence-backed, revisable skills without retraining the base model. It matters because it replaces raw trajectory retrieval with governed skill promotion, validation, and retirement.
- Separates memory into trace evidence, reusable policies, and environmental knowledge; only stable, beneficial policies become callable skills.
- Outperformed or matched baselines across five EvoAgentBench domains, with the largest success-rate gain in software engineering at 15.39 percentage points.
- Ablations showed flat memory and removing skill consolidation substantially reduced Pass@1, supporting structured experience management over simple vector retrieval.
- Limitations include heuristic—not causal—credit assignment, reliance on prompted LLM operators, runtime-specific evaluation, and unresolved privacy and security risks.
Sources (1)
做过十遍还要从头摸索？Agent 的记忆终于开始长成技能
TL;DR - MSCE is an arXiv framework that turns long-horizon agents’ accumulated experience into evidence-backed, revisable skills without retraining the base model. It matters because it replaces raw trajectory retrieval with governed skill promotion, validation, and retirement.
- Separates memory into trace evidence, reusable policies, and environmental knowledge; only stable, beneficial policies become callable skills.
- Outperformed or matched baselines across five EvoAgentBench domains, with the largest success-rate gain in software engineering at 15.39 percentage points.
- Ablations showed flat memory and removing skill consolidation substantially reduced Pass@1, supporting structured experience management over simple vector retrieval.
- Limitations include heuristic—not causal—credit assignment, reliance on prompted LLM operators, runtime-specific evaluation, and unresolved privacy and security risks.