Lab
A good memory must know when to retrieve nothing
Short answer
Reliable memory should not force a match for every question. When nothing is relevant enough, abstaining is better than presenting a false precedent with confidence.

Most memory demonstrations show the same thing: a question, then a memory that gets retrieved.
That is useful, but incomplete.
One of the hardest cases appears when nothing should be retrieved at all.
Why weak memory always finds something
Semantic search works through similarity.
Two pieces of text can use related concepts without referring to the same decision.
A question about a new supplier can look similar to an older logistics problem. A new strategy can resemble an old brainstorming note. A topic that never existed can still have a nearest neighbor.
If the system must always return something, similarity eventually gets mistaken for precedent.
An AI can then build an excellent answer around the wrong memory.
“Retrieved” does not mean “relevant”
Memory should distinguish at least three situations:
- genuinely relevant context was found;
- there is a possible connection, but it is uncertain;
- nothing is reliable enough to influence the answer.
That third outcome matters.
It prevents a weak match from acquiring the authority of real history.
The problem grows with memory
With a few memories, topics are often easy to separate.
After months of work, the same words recur, decisions resemble each other, people change their minds, and similar problems happen in different contexts.
The challenge is no longer only retrieval.
It is also rejection.
A concrete example
You ask: “Do we keep the same provider for the website?” The memory holds a long discussion from a year ago about another provider, for hosting. The topic is close, the decision is not the same.
A memory that forces a match brings that discussion back, and the AI answers as if it applied to the website. A careful memory lets the AI notice that no decision about the website exists yet. You lose two seconds, and you avoid an answer built on a false precedent.
The user outcome
For the user, that caution changes the final answer.
Without it, AI may say:
“You had already decided that…”
when it only found something vaguely similar.
With more careful memory, it can say:
“I cannot find a precedent reliable enough to use your history as a basis here.”
That sentence sounds less impressive.
It is much more useful.
Trust also comes from abstention
Memory does not become trustworthy because it always finds something.
It becomes trustworthy when users can understand the difference between:
- context found;
- possible connection;
- no reliable context.
Knowing when to stay silent is part of precision.
The mechanism is detailed in when semantic search invents a false precedent. We also test topics that do not exist, because retrieving is not verifying: see why found does not mean true. For the effects of scale, read limits of retrieval at scale, and for our studies, the Lab.
Limits of abstention
Abstaining is a balance. A memory that is too careful misses useful context, and one that is too sure of itself invents precedents. No setting removes both errors completely: we measure them rather than pretend they are solved.


