The authors formalize agent memory as a control problem and propose REM, which consolidates each completed trajectory into a phase-structured document (localize, understand, fix, verify) that is two orders of magnitude smaller than the raw trace.
Governing memory lifecycle, not just retrieval, boosts LLM agent reuse
REM improves bug-fixing resolution from 24.9% to 27.6% at 17% lower cost by consolidating trajectories, gating recall on agent struggle, and evicting after one use.
Big Tech
Xinyi Lisa Chen · Pedro Herrero-Vidal · Xiaohan Chen · Xiaojing Yu · Cris Liu · Bingqing Song · +3 more
Amazon · Yale University
Research Digest··2 min read
The authors introduce REM, a framework that manages experience memory inside an LLM agent’s think-act loop through three explicit policies: what to store (a phase-structured document), when to recall (withholding implementation details until the agent struggles), and what to forget (eviction after single exposure).
§