Write-access controls and integrity validation on memory

Control objective

This control governs who and what may write to an AI agent's persistent memory and proves that stored memory has not been tampered with — a foundational safeguard because poisoned or silently altered memory (MITRE AML.T0080.001, OWASP ASI06, MAESTRO L2) can steer an agent's future decisions, for example planting a false 'pre-approved' flag in a customer's stored profile. It checks that every memory write is properly authorised and that each stored record carries an integrity hash so later reads can detect modification. The targets are explicit: 100% of memory writes authorised and hashed, and a mean-time-to-resolve for any anomalous-write alert of under 24 hours. To implement, place an authorisation gate in front of the memory store that validates the writer's identity and entitlement, compute and persist a cryptographic hash with each record, and verify that hash on read; log writer identity, timestamp, hash and authorisation decision as evidence. Add anomaly detection for unusual write patterns (volume spikes, off-hours writes, writes to protected fields). A breach — an unauthorised or unhashed write, or a hash mismatch — must trigger an anomalous-write alert that is investigated and remediated within the 24-hour MTTR window, with the suspect memory frozen or rolled back until cleared.