Enforcing the Enforcement Layer: Bidirectional Governance Integrity Under Agent Amnesia

This paper introduces bidirectional governance integrity (M2), a repository-governance property for AI agents operating across fresh sessions, context compaction, and other forms of agent amnesia. M2 protects against two complementary failures: required enforcement members silently disappearing and previously forbidden members silently returning. The proposed design uses a canonical manifest and an environmentally invoked, fail-closed reference monitor to reconcile the live enforcement set against declared requirements. A deterministic demonstration shows that this set-level approach detects required-member absence that object-level checking misses. The contribution is a focused transfer and composition of established security mechanisms for AI-maintained repositories, with explicit trust boundaries and limitations.