Governing LLM Agent Memory via FC and PS Admission
October 9, 2026
Two inference-time methods, factor-compiled (FC) and permission-semantic (PS) admission, manage persistent memory to prevent sycophancy and cross-domain leakage. These approaches reduce judge-assessed failure rates by up to 8.8 percentage points without requiring model retraining.
HOW THIS AFFECTS YOU
●
builderYou can implement these admission policies to improve agent personalization without retraining.
●
researcherThese methods provide a framework for decoupling memory retrieval from context injection.