Persistent Conversational Identity as First-Class Runtime State

Long-running conversational agents can fail in a way users notice even when individual answers remain technically competent: the agent becomes less recognizably itself. Persona fidelity weakens, generic assistant cadence reappears, stale interaction rules resurface, and accumulated history begins to blur the distinction between who the agent is, what happened before, and what matters only for the current task. Recent long-dialogue evaluations document persona-fidelity degradation over extended interaction, particularly when models must sustain both task performance and role consistency. We argue that part of this problem is architectural. Persistent conversational identity is usually represented as prompt text, retrieved memory, accumulated chat history, or some combination of the three. None of these representations gives identity its own lifecycle. We instead treat persistent conversational identity as first-class managed runtime state: an explicit state object with its own representation, version, loading semantics, invalidation conditions, transfer boundary, and verification surface. This abstraction separates four state classes that are frequently collapsed in prompt-centric systems: construction source, runtime identity, continuity state, and turn-local context. The separation yields three design consequences. First, a new user turn is not an identity revision. Second, new history does not automatically become identity. Third, the representation of identity is itself behaviorally consequential because a generative model consumes state as conditioning, not merely as deterministic configuration. The strongest representation-level evidence came from a retained ablation in which identity dimensions such as relationship_address, self_role_label, and signature_mannerism remained explicitly present with state = unbound and fallback = omit. Under high-expression isolation, unsupported identity-surface completion still appeared. A redesign that removed irrelevant ontology slots entirely—strict omission—avoided the motivating failure across six fresh-isolation runs spanning three different Move classes. The result motivates a general principle for LLM systems: explicit absence and non-representation cannot be assumed to be behaviorally equivalent. To show that the state abstraction is implementable, we present the Compiled Resident Projection (CRP) as a reference runtime architecture. CRP compiles stable Resident semantics into a reusable artifact, recompiles on source change rather than user turns, supports explicit binding and fail-closed loading, separates construction material from runtime state, and permits bounded migration with continuity kept distinct. A separate preregistered six-session sanity check compared a semantically matched monolithic prompt and a managed runtime on GPT-5.6 Sol High. Both reached ceiling on the 18-turn battery, establishing short-horizon behavioral parity rather than superiority. The contribution is therefore not “a better persona prompt.” It is a different systems abstraction: persistent identity should be managed as persistent state. CRP is one implementation of that abstraction; broader questions of control, ownership, and behavior remain outside the scope of this paper.