Operative Contexts: Belief Revision and Memory in Agentic AI

Long-term memory is becoming central to personalized and agentic AI systems, yet it is often evaluated primarily as retrieval: can the system recall past information? We argue that aligned long-term memory requires a deeper form of control. Agents do not merely store user information; they construct an operative context through which future requests are interpreted and acted upon. Drawing on common ground, cognitive control, and belief revision, we distinguish stored memory, context-relevant information, and the information that actually guides behavior. We define silent contextual misalignment as a failure in which outdated, uncertain, wrongly scoped, or contextually inappropriate information guides action while remaining hidden from the user. We propose desiderata and diagnostic stress tests for memory update, scope, task demand, and uncertainty control, and argue for contextual inspectability as a mechanism for collaborative repair. Persistent memory should therefore be evaluated not only for what an agent can retrieve, but for whether it selects, revises, and exposes the right context for the user's current task.