Summarization and compaction are emerging as a reliability hazard, not just an efficiency win: compressing history systematically trades away factual fidelity and can fabricate policy.
Evidence
- Compaction vs RAG fails precisely when verbatim phrasing is needed, despite token savings
- ‘When Do Summaries Become Instructions’ shows self-written handoff summaries hallucinate policies/permissions that drive downstream errors
- ProcessPrefBench finds injecting historical preferences improves alignment but doubles factual fabrication
- Task-Conditioned Utility work shows task-blind summarization often degrades memory utility versus raw context
Implications
- Provenance tagging of requirements and task-conditioned evaluation become necessary mitigations before deploying memory compaction
- Evaluation must weigh hallucination/fabrication risk against token savings rather than optimizing compression alone
Concepts
Agent Memory Context Engineering LLM Evals RAG
Confidence
high