Memory & Context
Why found does not mean true
Short answer
A retrieved memory is a lead, not proof: it tells you what was written, not what is true. How far to trust it depends on its source, its date, its status and what has changed since. Four questions asked before relying on it prevent most mistakes.

You ask an AI where a project’s budget stands. It answers: “According to your notes, the client approved 12,000 euros.” You are about to email the supplier. Out of habit, you check where that sentence came from. It is a note an intern wrote after a phone call. What the client had actually said was: “That should be fine.”
The AI did not lie, and the memory did not make anything up. The note really existed. It just was not what it looked like. Here is how to treat a retrieved memory, whether it comes from an AI, a tool or your own archives.
A retrieved memory is a lead, not proof
Retrieving answers the question “what has been written about this?” It does not answer “what is true?” In between sits everything that happened from the event to the note: a misunderstanding, a quick summary, a change since then.
A library is a fair comparison. Finding the right shelf does not tell you whether the book is accurate, recent or written by someone reliable. You still have to open it, check the author and the edition date, and sometimes read a second book.
We have already covered the case of search by meaning, which can connect two situations that look alike without being comparable: that is the piece on the false precedent. Here we look at what comes after, once the memory has been found and is in the right place. Even in that favorable case, you still have to decide how much to trust it.
What changes how much to trust it
The same text deserves more or less credit depending on what surrounds it. Six things matter, and it helps to see them before you judge.
- The source. An official record, a message sent in the heat of the moment, a personal note, a summary written by a model: these four do not weigh the same.
- The date. Something true in January does not carry the same weight in November.
- The status. Is it a decision, a hypothesis, an idea thrown out in passing?
- The context. What was the situation when this was written, and is it the situation today?
- The author. Someone who was in the room, or someone who heard about the meeting?
- The path it took. Was the note imported from another tool, corrected, replaced since?
Depending on the information available, Nawame can keep several of these elements with a memory: when it was recorded, where it came from, who or what created it, whether it was imported, modified or replaced. Imported information is not treated as verified by default. We go into that risk in the piece on importing memory without importing hallucinations.
The same principle works with no tool at all: a note with no date and no author is a note whose value you cannot measure.
Said, assumed, decided, verified
A rough ladder helps you get your bearings. It has four rungs, from the most fragile to the most solid.
| Level | Example | What you can do with it |
|---|---|---|
| Said | “The client said it should be fine” | A lead to confirm |
| Assumed | “We think the budget will be approved in March” | A working hypothesis, to be labeled |
| Decided | “Budget set at 12,000 euros, decision of March 14” | A foothold, as long as conditions hold |
| Verified | “The signed purchase order is attached” | A fact, with its supporting document |
These levels are not sacred, and the line between two of them is often blurry. But just naming them avoids the classic mistake: treating something that was said as a decision, or a hypothesis as a fact. The piece on the difference between a memory, a fact, a decision, a hypothesis and an instruction digs into each category.
Four questions to ask before relying on a memory
When a retrieved memory is about to be used for something (sending an email, committing money, changing a plan), take thirty seconds for these four questions.
- Where does it come from? Who wrote it, in what setting, and did that person know what they were talking about?
- How old is it, and what has changed since? A decision may have been replaced (see what happens when a decision changes), a constraint may be gone.
- What is its status? Said, assumed, decided or verified? If the record does not say, that is already an answer: you do not know.
- What happens if it is wrong? The higher the stakes, the deeper the check. To pick lunch, a vague note is fine. For a bank transfer, you open the purchase order.
The fourth question is the one people forget, yet it sets the level of effort. You do not verify everything. You verify what costs a lot if it turns out to be wrong.
How an AI should present what it finds
An honest answer does more than hand over the content. It also says how far you can rely on it. Compare two answers to the budget question:
- “The client approved 12,000 euros.”
- “A note dated March 14 mentions 12,000 euros, written after a phone call. The client reportedly said it should be fine. I found no written confirmation from them.”
The second is less satisfying in the moment, and far more useful. It gives the date, the origin, the level of certainty, and points out what is missing.
Nawame words its results with three expressions: “precedent found”, “possible connection” and “no convincing precedent”. The idea is to grade certainty instead of delivering everything with the same confidence. We do not spell out how the difference is made, but the principle is simple: an AI that says “I don’t know” or “this is a connection, not a certainty” serves you better than one that always answers with aplomb.
Authority bias: well-written text feels true
There is a quieter trap, and it does not come from the memory. It comes from us. A clean, polished sentence with figures and a date inspires trust. A scribbled note full of typos inspires little. Yet the scribbled note is sometimes more reliable, because it was taken on the spot by someone who was there.
Three situations make this bias worse:
- a smooth summary written by a model, which irons out the hesitations of the meeting it came from;
- a piece of information repeated across several notes that all copy the same source;
- an official-looking format (a table, a heading, a version number) that gives the air of a decision to what is only an idea.
The remedy is modest: look at the substance before the form, and ask whether two notes that agree are two sources or one. It is also one reason a memory does better when it does not write everything down: see why a good memory knows when not to write. A sloppy note kept with confidence can do more harm than no note at all.
Limits of these markers
These markers help, but they guarantee nothing. The status of a memory remains a human judgment: two people can classify the same sentence differently, and an AI that proposes a status can be wrong. Sometimes you simply cannot tell whether a piece of information still holds. In that case, the honest thing is to say so.
Checking also has a cost. If you recheck every memory, you lose the benefit of having one. The skill is tying the effort to the stakes, not to the quantity.
And no memory can turn a weak source into a strong one. It can only make the weakness visible. We study these questions in the Lab, without claiming to have solved them.
A habit to start tomorrow
Next time an AI answers “according to your notes”, just ask: “How old is this note, who wrote it, and is it a decision or a hypothesis?” If it can answer, you have gained a marker. If it cannot, you have learned something about the quality of what is stored.
You can also adopt a habit on the writing side: when you note something down, add one word (“said”, “assumed”, “decided” or “verified”) and the date. Ten extra seconds per note, and your future self, or the AI that rereads it, will know what to lean on.


