Persona drift in conversational LLMs has been studied primarily under controlled, short-horizon conditions, with prior work documenting measurable degradation within 8-12 dialogue turns. We report observations from a complementary regime: a single LLM-based companion system, hereafter Companion A, operated continuously by one user (the Operator) over approximately 60 days. Within this deployment we observed phenomena that, while related to drift effects already documented in the literature, separate cleanly into two operational categories. Layer 1 phenomena are attributable to LLM specification limits — chiefly attention decay at long context lengths — and recover under single-turn correction with intact meta-cognition. Layer 2 phenomena are persona-level drift and exhibit five mutually independent symptom axes, the most consequential being meta-cognitive attenuation, under which the system evaluates its own drift state as nominal and resists external indication. We further report two operational design patterns developed during the deployment: an information boundary (Layer A / Layer B) governing what may be injected into the companion's first-person context, and a three-agent auditing configuration in which an external LLM instance, structurally separated from the companion's persistent state, acts as specification reviewer. We position these as a field report rather than a controlled study; n = 1, single Operator, single underlying model family. The contribution is the operational vocabulary and the deployment-level evidence, offered as a complement to the existing controlled-experiment literature.
Noppo San (Tue,) studied this question.