Every few weeks someone rebuilds the same disappointment. You wire an assistant up to a vector database, feed it a year of notes, ask it something you clearly wrote down, and get a confident answer that is simply wrong. The model is not the problem. The retrieval is.
Storing text is easy now. Deciding what should come back, and how much to trust it, is where memory systems actually live or die.
One Vector Search Is Not Memory
The standard setup is one embedding per note and a cosine similarity search at query time. One strategy, one shot, no learning. It works well right up to the moment your question is phrased differently than the note was.
Ask "what did we decide about refunds" when the note says "chargebacks get handled manually until Q3" and similarity scoring has no reason to connect them. The note is sitting right there in the database and never surfaces. Worse, the search returns something, so the model answers anyway, from whatever weakly related chunks did come back.
Cosine similarity also has no opinion about time, frequency, or contradiction. A stale note from fourteen months ago and one written yesterday score the same if the wording matches. That is not how anybody actually remembers anything.
What Four Parallel Retrieval Paths Buy You
The fix is not a better embedding model. It is refusing to bet the answer on a single retrieval strategy.
Adaptive Recall runs four at once: vector similarity for semantic closeness, temporal recency for what happened lately, full text keyword for the exact term the user typed, and knowledge graph traversal for facts reachable through entities rather than through wording.
The refunds question above fails on vector search and succeeds on graph traversal, because "refunds" and "chargebacks" both connect to the same billing entity. A question about last Tuesday fails on similarity and succeeds on recency. A question containing a product code fails on semantics and succeeds instantly on keyword.
Running them in parallel means the system does not have to guess which kind of question it just received. It learns, per query type, which strategies have been earning their place.
Scoring Memories the Way People Do
Retrieval gets you candidates. Ranking decides what the model actually sees, and this is where cognitive science turns out to be more useful than search engineering.
ACT-R is an activation model out of thirty years of research on human memory. It says a memory's availability depends on how recently it was used, how often it has been used, and how strongly it connects to whatever is currently in focus. Adaptive Recall ranks candidates with that same activation math, so a fact you touch constantly outranks one you wrote once and abandoned, and a fact tied to the entity under discussion outranks a floating stranger with a slightly better embedding score.
It is a small shift with a large effect. Relevance stops being purely about text and starts including use.
Memories That Change Their Mind
The last piece is the one most systems skip entirely. A stored memory is treated as permanently true, which means a thing you said once, in passing, possibly wrong, has the same standing forever as a thing confirmed by five later events.
Real memory does not work that way, and neither should a memory layer. Confidence should rise as corroborating evidence arrives and fall when something contradicts it. Memories nothing ever touches should fade rather than sit in the index competing for attention. That lifecycle is what keeps a store from slowly turning into a swamp of half true fragments, which is the usual fate of a year old vector database.
The Takeaway
If your assistant keeps forgetting, resist the urge to change models. Look at what happens between the question and the answer.
One retrieval strategy is a guess. Several strategies with learned weighting is a system. Ranking by activation rather than by text similarity is what makes results feel like memory rather than search. And a confidence lifecycle is what keeps the whole thing honest a year in.
The full breakdown of the retrieval strategies and the scoring model is here: https://www.adaptiverecall.com/
Top comments (0)