When AI Remembers You It Also Needs to Know When You Have Changed

Last week, you asked an AI assistant to keep an email short. This week, you need a detailed report. It trims the explanation anyway, because it has remembered that you prefer concise writing.

The missing word is “this.” A request for this particular task has quietly become a rule for every task that follows.

Anthropic reported on November 6, 2026: wiki-assisted search met 61% of grading criteria across 25 Stainless-document questions, versus 53% for search alone. Separately, Sonnet 3.7 and Fable 5 kept 413 versus 10 files in Pokémon, earning one versus four badges. Different models and single runs prevent isolating memory’s causal effect.

Those findings are a useful starting point for a practical question. Once an assistant can carry information from one task to another, how should we decide whether its memories deserve our trust?

A temporary choice can turn into a permanent assumption

Imagine booking an early flight because you have a meeting that afternoon. An assistant might save “prefers morning flights.” Months later, while planning a holiday, it keeps offering departures before sunrise.

The original choice was perfectly sensible. The trouble came when a summary dropped the reason for it.

The same problem can affect a work project. “The client is postponing this until the next budget cycle” becomes “the client is not interested.” Instructions for an older software release become advice for the current version. A tentative explanation of a bug becomes the assumed cause of every similar failure.

For a useful memory, I would want a few ordinary details kept together: who supplied the information, when it applied, what conditions mattered, and whether anyone actually confirmed it.

That may make a note slightly longer. It can also prevent a much longer conversation about why the assistant keeps making the same wrong assumption.

How AI organizes memory in the background

Anthropic uses “dreaming” as a metaphor for organizing AI memory in the background, rather than human dreaming or AI hallucinations. Its Dreams documentation describes a research-preview service that checks existing memory against past sessions and creates a separate, reorganized memory store. The inputs stay unchanged, allowing the output to be reviewed before it is adopted.

Think of it as revising handover notes after a piece of work. You have more perspective than you did while rushing to finish the task. You can see which decisions held up and which details need another look.

The important design question is what happens when that revision is wrong.

Suppose a meeting date was misrecorded in the first place. A beautifully rewritten summary can make the error look more settled. Better prose offers no protection against a bad premise.

For a product built around this approach, I would want a visible history of changes and a way back to the evidence behind an important claim. If the evidence remains ambiguous, the system should be able to leave the matter unresolved. A neat paragraph is a poor reason to turn uncertainty into certainty.

What work should a memory actually save

A wiki is a collection of linked pages, much like an internal reference handbook that can be updated over time.

Consider a project whose requirements live in a document, whose schedule lives in a spreadsheet, and whose latest decisions live in chat. Answering “Can we launch this week?” involves finding the relevant versions and checking how they fit together.

A maintained project note could give the next person or assistant a useful starting point: which release is under discussion, who owns the decision, and which condition remains unmet.

Some details still deserve a trip back to the source. A payment deadline, an exception in a contract, or a technical constraint that applies only to one release may lose its meaning when compressed into a friendly summary.

I would judge the note by the work it removes. Does it reduce the time spent locating evidence? Does it make the outstanding question obvious? Can someone reach the underlying document without another round of searching?

If all it does is repeat material in a new place, it has also created another place that must be kept current. That maintenance work belongs in the cost of the feature.

A small test set can reveal a lot

Teams considering an AI memory feature can learn a great deal from a few deliberately awkward scenarios.

First, change a fact. Give the assistant an initial launch date, then clearly replace it with a new one. Ask it to make a plan that depends on the date.

Next, make a one-time request. Ask for a formal email, then switch to a casual message. Check whether the earlier instruction has escaped its original scope.

Try two projects with similar names but different requirements. See whether information from one leaks into the other.

Finally, ask for a memory to be removed. Return to the subject with different wording and check whether the old assumption still shapes the answer.

These are proposed tests, not experiments performed for this article. To make the comparison useful, hold the model, available information, and task difficulty steady. Repeat the tasks and record correctness, time, and cost. Changing several things at once makes it difficult to tell which change helped.

That approach is less glamorous than a feature demonstration. It is also much closer to the mistakes that make an assistant frustrating to use every day.

People need a way to correct the record

A memory feature deserves more than an on-off switch.

Users should be able to see when an assistant has turned a passing comment into a lasting preference. They should be able to narrow it to one task, update it, or remove it. When a project ends, its assumptions should not keep resurfacing in unrelated work.

Remembering also needs a clear boundary from permission. Knowing which coffee someone usually buys does not settle whether the assistant may place and pay for an order. A product should handle those as separate decisions.

For most people, the best test of an assistant’s memory will be straightforward. Does it accept a correction? Does it understand the difference between “today” and “usually”? When circumstances change, does it check before relying on an old conclusion?

A useful assistant saves you from repeating the background. A trustworthy one also saves you from having to argue with a version of yourself it recorded months ago.