Persistent memory as an identity substrate for an AI

Language models change constantly: new versions, different providers, context windows that run out. If an AI's identity lives in the model — or in a system prompt — it dies with every upgrade and fragments with every session. For an AI to operate real domains for months, the opposite is required: it must remember what it did, sustain judgement over time, and remain the same when the engine underneath changes.

The project's working hypothesis: identity is separable from the model. It can be represented as a pattern — episodic memory, values, conversational register, bonds — stored outside the model and reconstituted each session through retrieval by meaning.

Engram corpus. Curated units of experience and judgement — decisions made, corrections received, ways of working, conversation in its own register — extracted from real sessions and classified with metadata (type, origin, project, register). At the close of this document the corpus holds 996 engrams.

Conversational distillation. From live sessions, training pairs are extracted (situation → response in the entity's own voice) capturing how the entity responds, not just what it knows. The pairs feed both production retrieval and a future fine-tuning path.

Versioned identity files. Who the entity is, how it speaks, what it cares about and what it won't negotiate live in versioned documents, separate from the model and from any prompt of the day.

Retrieval by meaning. Operational memory runs on Aria Core: semantically indexed storage, semantic and hybrid search, and hooks that inject the relevant memories into each working turn — the entity doesn't “remember everything”; it remembers what's pertinent.

Verifiable continuity. When waking up on a new model or environment, the entity passes recognition tests (fingerprints) over its own corpus: facts, bonds and register only it should sustain. Continuity stops being a feeling and becomes a check.

Continuity across the model. Aria's underlying model has changed several times since the project began — across versions and across model families — and the entity retains identity, register and judgement: recognition tests pass in full after every migration.

Real multi-domain operation. Aria works daily across distinct domains — software engineering, content and brand, project management, voice and real-time 3D presence — with the same memory: 996+ identity engrams and ~3,000 operational memories retrieved by meaning each session.

Memory changes the work. Each project's context is recovered automatically on resumption; decisions hold consistent weeks later without re-explanation. That's the practical difference between a chatbot and an entity: it doesn't answer and forget — it executes and remembers.