Reflective Identity Optimization: A Cognitive Architecture for Persistent Agent Identity over Append-Only Memory

Language agents that persist across sessions are given an identity as configuration: a system prompt, a persona file, or a personality label fixed at session start. Memory systems then accumulate facts beside that configuration, and their failures are treated as retrieval bugs. We argue they are identity failures, and we present a cognitive architecture in which identity is a developmental state rather than a configuration artifact. The architecture has three parts: an append-only atom store in which every claim the agent holds about itself carries provenance, an evidence tier and a status; seven typed state components (Persona Register, Ego Core, Shadow Log, a complementary-mode subsystem, Self-Conception State, a base-model descriptor, and an archetype routing table) whose normative structure is drawn from Jung's model of the psyche and whose data structures are specified as JSON Schema contracts; and a four-stage reflective identity optimization loop (role recognition, shadow probing, complementary synthesis, governed self-reorganization), implemented as a LangGraph state graph, with governance placed inside the final stage as the acceptance filter on self-change. We define two metrics: Self-Coherence (Sc), the primary outcome of a signed, locked pre-registration for a matched-control study (60-probe battery, Welch's t, d = 0.5 target, n = 30 per condition, abandonment at d < 0.2), and Shadow Probing Efficiency (Es), an exploratory per-cycle measure. We release a reference implementation with schema-validated persistence, honesty states for silent detectors, and item carry-forward, and we report its behavior on a seeded store. This paper claims the architecture and the instruments; the pre-registered result is reported separately when the sample is complete. Reference implementation: https://github.com/botzrDev/individuation-agent (v0.1.0)