Most AI products fake personalization. We built the version that actually remembers.

Most AI products fake personalization by re-reading the last few turns of chat history. The agent sounds personal for the length of a session and forgets you the moment a new one starts. We build the version that actually accumulates context over months — that distinguishes what the user told it from what it merely guessed, lets the user see and edit what's stored, and refuses to pretend it knows things it doesn't.

01 · Two classes of memoryWhat the user said vs. what the agent guessed

Every entry in the profile carries a source. Confirmed entries come from explicit user statements — "from now on, always cite sources" — and are exempt from decay, period. Inferred entries come from session analysis and require at least two reinforcements before the agent treats them as established enough to surface in its prompt. A single-pass observation can be written, but it sits below the surface until reality confirms it again. The two classes never blur: an inferred update can never silently overwrite a confirmed entry — the store returns a conflict and waits for human approval.

02 · Commitments vs. ventingWhat turns into a memory — and what doesn't

The difference between "I'm done with TypeScript" and "From now on, use JavaScript instead of TypeScript for all my code" is the difference between venting and a directive. The first is frustration; the second is a commitment. Most products treat them the same and end up with prompts full of half-meant complaints. A fast-tier LLM watches every turn for the directive shape — explicit cues like from now on, always, never — and rejects hypotheticals, frustration, and idle musing. Only committed directives become confirmed memories.

03 · The association graphFacts that cluster reinforce each other

Profile entries don't sit in a flat list — they live in an association graph that links facts in the same category, sharing tags or vocabulary. Tightly-clustered facts reinforce each other: a recently-retrieved entry lifts the relevance score of everything it connects to. Isolated facts — the one-liner you said six months ago that never came up again — drift below the relevance threshold and get archived. The decay score is a weighted composite of confidence, recency, graph connectivity, and reinforcement count; entries below 0.2 are archived, below 0.05 are deleted. Confirmed entries are explicitly excluded from the sweep, every time.

04 · Cold-start honestyThe product refuses to pretend

The most damaging thing a personalized agent can do is fake familiarity on the first session. When the profile has fewer than five entries, the retriever returns an empty result — explicitly, not a half-confident guess. The system prompt receives a fallback string that tells the agent to ask clarifying questions instead of acting as if it remembers things it does not. It's the same instinct as the confidence tag on retrieval: the architecture knows when it doesn't know, and says so.

05 · User-facing controlsAuditable, not magic

The user sees exactly what the agent remembers. Inferred entries can be confirmed (promote to immortal), dismissed (delete the suggestion), or edited. Any entry can be force-forgotten. The whole profile can be re-derived from raw interaction logs — a fresh pass that rebuilds the inferred portion from history without bias from the existing entries. The whole thing can be exported as JSON. There are ten profile endpoints exposed for this; nothing about the memory layer is hidden from the person it's modeling.

06 · Why we built it this wayThe memory layer enterprise buyers ask for

Most teams ship "memory" by stuffing the last N turns of chat history into the next prompt and calling it personalization. That works for a session. It does not work for a relationship. Once a customer is six weeks in and expects the agent to know them, the cracks show: the model forgets stated preferences, invents ones the user never said, and has no way to tell the user what's stored. We built this engine because every serious deployment we've shipped needed it — separation of confirmed from inferred, a filter that rejects venting, a graph that decays isolated facts, a cold-start that admits ignorance, and a panel where the user can see and edit everything.

None of these pieces are exotic on their own. The work is in stacking them so the agent stops being a goldfish in a polite tone of voice and starts being something a customer can actually live with for a year.