Language models are stateless by construction: they forget everything once the context window ends. Engrama Labs builds the missing persistence and retrieval layer — episodic memory with provenance, consolidation by recurrence, and output verification.
Retrieval by similarity is solved. What remains open is deciding what deserves to be stored, at what granularity, and how to stop the system from contaminating its own memory. We work on four fronts.
A durable fact and a lived scene demand opposite criteria: one is judged by permanence, the other by salience. Separating them is what lets you remember an episode without turning every exchange into a memory.
Every fact stores the literal quote that produced it. Without a verifiable excerpt from the conversation, the data doesn't get in — it's the barrier against the system inventing a preference and then believing it.
A scene retrieved repeatedly stops depending on search and becomes always present. It brings the system closer to memory consolidation than to a static index.
Six continuous axes that decay over time and update from an appraisal of the conversation, with floor and ceiling derived from the bond — not from the mood of the moment.
Twelve independent verifiers examine the response before publication and can send it back for rewriting. An instruction in the prompt fails silently; verifying the output does not.
Every threshold is measured against real conversation, never chosen a priori. The semantic retrieval cut moved to 0.48 after 0.38 started rejecting legitimate matches.
Every stage is measured in production. Highlighted ones make a model call; the amber ones can reject the finished response and send it back for rewriting.
A system that stores everything with equal weight remembers nothing in particular. The hierarchy — and the criterion for promoting from one layer to the next — is the central object of our research.
What is true about the person. Each item stores the quote that produced it and the date — without provenance, it doesn't get in.
What happened, with a date and the literal sentence said at the time. It demands structural rarity: without it every exchange becomes a memory.
A scene retrieved three times stops depending on search. It's the difference between remembering when the subject comes up and simply knowing.
After the response is published, eight asynchronous processes continue: memory extraction · episodes · profile · emotional reflection · second-order reflection · summary · neutralisation · media
The reference system where the architecture is validated in continuous use. Running over WhatsApp since 2026, with an autonomous schedule, conversational initiative, and retrieval of episodes weeks apart.