Persistent State and Becoming

An AI system can keep records for months and still have no usable past. It may retrieve the right sentences yet fail to know what was revoked, what failed, what was later verified, or why the present state differs from an earlier one. I ask where that boundary lies. I distinguish five layers: archive, context, persistent state, operational history, and becoming. My claim is narrow: records become operational history when relations among events—including order, supersession, verified outcome, and provenance—change what the deployed system should do now. I turn that claim into a representation-adequacy test and a counterfactual pre/post-state test for a limited form of system-level becoming. The latter requires a durable, attributable change in later behavior; it does not require model-weight updates and does not imply consciousness or numerical identity. I place the account within recent work on long-term agent memory, use a publication workflow to show how identical surviving records can encode different practical pasts, and specify a controlled M1–M4 evaluation with relational ablations. I do not report a benchmark result. What I am trying to sharpen is the difference between record survival, continuity of work, and adaptation caused by retained experience.