Operative Contexts: Belief Revision and Memory in Agentic AI

Persistent memory is becoming central to agentic AI systems, yet memory evaluation remains largely focused on retrieval: can the system recall past information? We argue that this framing misses a distinct problem of context control. Agents must determine which remembered information is relevant to the current task, which information is uncertain or outdated, and which conflicts should be surfaced before action. Drawing on work in common ground, cognitive control, and belief revision, we distinguish stored memory from operative context: the information that actually guides an agent’s behavior. We define silent contextual misalignment as a failure in which an agent acts from an inappropriate operative context while this mismatch remains hidden from the user. We propose desiderata and diagnostic stress tests for evaluating context-sensitive memory control in agentic systems, including conflict revision, scope control, task-demand selection, and uncertainty preservation. More broadly, we argue that reliable agent memory requires not only retention, but mechanisms for selecting and exposing action-guiding context.