acceptodds
Desk rejected at ICLR 2027

Bridging Reflective Memory and Procedural Skills: Context-Anchored Policy Compilation for LLM Agents

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Reflective memory systems for LLM agents preserve execution fidelity but incur substantial context overhead. Compressing reflective memories into reusable skills promises efficiency, yet recent benchmarks show distilled skills often underperform their source memories. Prior work attributes this to information loss during compression. We hypothesize that the core issue is context-detachment: procedural knowledge loses effectiveness when stored as global guidelines rather than anchored to the skill components that trigger it. To address this, we propose Skill-Local Policy, which anchors compiled procedures to relevant skill subcomponents rather than global guidelines. Controlled ablation on six SkillFlow task families confirms our hypothesis—identical content achieves 58.3% success as flat guidelines versus 83.3% when context-anchored. Large-scale evaluation on 347 matched pairs shows context-anchored compilation provides safety (zero within-family regressions) and modest effectiveness (+3.8pp overall), with benefits concentrated where procedural structure aligns with skill decomposition. Code and data are available at \url{https://anonymous.4open.science/r/skills-policy-C1BE

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
LLM agents, agent memory, skill learning, procedural knowledge

Related papers

Discussion (0)

Sign in to comment.