MemSentry gates persistent memory writes on a signed security-state delta rather than on content classification

arXiv·medium signalXBlueskyLinkedInCopy link

Memory poisoning lets adversarial content stored in long-term memory suppress security alerts, enable privilege escalation, alter trust relationships or override policy without touching model weights or system prompts. MemSentry intercepts proposed persistent-memory writes and returns deterministic Accept, Review or Quarantine decisions, scoring each write on source trust, semantic risk, attack radius over a component-dependency DAG, access risk, and a signed delta capturing whether the operation weakens or strengthens security posture, instantiated on a 20-asset dependency DAG with a 10x20 access matrix. Pairing the write-side gate here with the retrieval-side revocation failure measured in the same 48 hours gives both halves of the agent-memory integrity problem.