What to Admit and How to Present: Governing Persistent Memory in LLM Agents

Key points

• We distinguish two governance decisions: admission, which determines what recalled information enters the working context, and presentation, which determines how admitted information is expressed.1
• We implement two inference-time designs without retraining: factor-compiled admission (FC), which assesses whole memory entries, and permission-semantic admission (PS), which decomposes entries into typed units; both translate adjudicated attributes into eligibility decisions via deterministic policies.1
• A query-conditioned gating baseline shows no significant change in objective-fact failure or pooled failure.1
• These results support evaluating admission and presentation separately: selection quality provides task-specific safety gains, while preserving beneficial memory use remains unresolved.1

Sources (1)

• [1]What to Admit and How to Present: Governing Persistent Memory in LLM Agents
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 03:45 AM

  Persistent memory can improve personalization in LLM agents but can also induce sycophancy and cross-domain leakage.

  We distinguish two governance decisions: admission, which determines what recalled information enters the working context, and presentation, which determines how admitted information is expressed.

Extractive summary: sentences quoted from the sources.