Persistent memory is often presented as a feature that makes AI more useful. An agent that remembers previous interactions, decisions, preferences, and events can operate with greater continuity and require less repeated instruction. But once memory begins influencing future decisions, memory becomes part of the governance architecture.
What an AI system remembers affects what it subsequently believes, recommends, and does. This raises questions that go well beyond storage. Who is allowed to write to memory? Which sources are sufficiently trustworthy to become persistent? How are observations distinguished from assumptions? What happens when stored information becomes outdated? How can incorrect or manipulated memories be identified and corrected? And which memories should expire rather than become part of the system's enduring view of the world?
These questions become particularly important because memory can transform temporary errors into persistent ones. A misleading input that affects one response is limited in scope. The same input incorporated into long-term memory may influence hundreds of future actions and may be passed from one agent to another.
Memory therefore needs its own integrity controls, provenance, permissions, validation, and lifecycle. The relevant question is not simply whether an AI system should remember. It is what deserves to become part of the world the system acts upon.