Privacy & Portability

Importing memory without importing hallucinations

Short answer

A memory exported from another assistant is a set of summaries and guesses written by a model, not a file of verified facts. To import it without importing its mistakes, sort it, date it, label it "imported, unverified", and promote by hand only what you have checked.

An open glass box with grey sheets coming out of it; each sheet passes under a small magnifying glass, and only a few get an amber thread that leads into a stone filing cabinet.

You are switching assistants, and the new one offers to import your old memory. The file arrives: thirty neat lines saying you work in logistics, prefer a casual tone, and are moving to Lyon in March on a budget of 450,000 euros. It all looks reasonable. Except that budget was never decided. You mentioned it one evening as a scenario.

That is the risk with importing. What comes out of an assistant is not a record of facts, it is a reconstruction. And a reconstruction pasted into a new place quickly turns into a reference, because it reads well and looks like it came from you. This article suggests a way to sort before you bring anything in.

Why an exported memory is not the truth

An assistant’s memory is almost always the output of some processing. According to OpenAI’s help center (read on September 30, 2026), ChatGPT’s personalization can draw on several sources: past conversations, memories saved at your request, custom instructions, files, and sometimes content from connected apps. In other words, what the tool “knows about you” is already a blend of things you said and things it worked out.

On the Claude side, Anthropic’s help center (read on September 30, 2026) says the assistant can import memory from other assistants and export its own, and describes this feature as experimental. We will not say more than that. The details move fast, and the official page is the reference.

Either way, the point is the same. A line of memory tells you “what”, rarely “where it came from” or “how sure we are”. When you import, you get statements whose status was lost on the way.

Fact, summary, inference: three different things

Take the moving line again. It could have three very different origins.

  • A fact. You wrote: “We are moving to Lyon in March.” Those are your words.
  • A summary. Ten messages covered neighborhoods and prices, and a model boiled them down to: “Moving to Lyon, budget 450,000 euros.” The summary compressed, rounded off, and maybe merged two ideas.
  • An inference. You asked about schools a few times, and the model concluded: “has school-age children.” You never wrote that.

All three arrive looking the same: a clean, confident sentence. That is what makes importing tricky. Our piece on the difference between a memory, a fact, a decision, a hypothesis and an instruction goes through these categories. For now, the point is that you need to tell them apart before filing them in the same place.

There is a second mix-up, about time. An exported memory describes a history, not the current situation. “Looking for an apartment” may have been true in January and be false by June. Without a date, you cannot tell.

What you should never promote automatically

Some kinds of entries deserve a line-by-line check, because a mistake there is either costly or easy to miss:

  1. Inferred preferences. “Likes short answers” may come from three rushed exchanges on a Tuesday.
  2. Dates. A one-month slip moves a deadline, and summaries get this wrong easily.
  3. Numbers. Budgets, quantities, prices: check them against the source, not against the summary.
  4. Decisions. An option that was considered is not an option that was chosen.
  5. Anything about other people. What a memory claims about a colleague or a client calls for extra care, and sometimes should not be imported at all.

What these share is that the mistake is plausible. A plausible mistake is the kind nobody notices. We come back to this in why found does not mean true.

A careful import, step by step

This works with any tool, including a plain document.

  1. Sort. Read the whole export once without fixing anything. Cross out what is trivial or out of date. Not writing starts here: not everything deserves to come through.
  2. Date. Give each line you keep an approximate date, even if it is “sometime last autumn”. A fuzzy date beats no date.
  3. Label. Tag the whole batch “imported, unverified”. That tag matters more than the content, because it tells you how to read everything else.
  4. Check a sample. Pick ten lines at random and look for their source: a conversation, a document, a reliable memory of your own. If more than two or three turn out wrong or impossible to trace, treat the whole batch with suspicion.
  5. Promote by hand. What you have checked moves to “confirmed”. The rest stays imported: searchable, but carrying no authority.
  6. Compare with what you already have. If you already keep notes on the subject, look for contradictions before merging anything.

None of this takes long, and it is the one place where quality gets decided. An unsorted import feels like a time saver for a week, then costs you a lot of time when you try to work out where a false statement came from.

What Nawame does on import

Nawame can allow you to import the memory of other AI tools. Importing is voluntary: nothing comes in unless you ask for it. Imported information is not treated as verified, and Nawame keeps imported content separate from stronger forms of authority, so that something reconstructed elsewhere does not become an established fact without anyone deciding it should. An import can also compare what is coming in with the memory already there and ask you to review possible overlaps.

This matches steps 3 and 6 above, but it does not replace the others. The details are in the imported memories section of the documentation.

Current limits

A summary written by a model is still a summary. Labeling a line “imported, unverified” does not make it more accurate, it only stops you from taking it on trust. Verification means going back to the source, and the source does not always exist. If a conversation was deleted, all that is left is the summary.

Keeping imported content apart has a cost too: someone has to decide, line by line, what deserves more authority. That is human work, and we do not claim to automate it. Finally, we have published nothing about the export formats of each third-party assistant, or about how faithful their summaries are. These features change quickly, and the vendors’ own pages are the reference.

A checklist before you import

Before you confirm an import, check five things:

  • I know where this file comes from and when it was produced.
  • I have read all of it, not just the top.
  • Dates, numbers, decisions and mentions of other people have been reread one by one.
  • The whole batch carries the label “imported, unverified”.
  • I know which wrong line would hurt me most, and I checked that one first.

If you are moving to a new assistant for one specific project, a hand-written brief is often enough, and we explain when importing is really worth it in how to move project context from one AI to another. And if the export itself is what interests you, what an AI memory export should contain lists what you can reasonably expect, while portability of meaning explains why a complete file is not always a usable one.

Your next prompt should not have to contain your whole story.

Start building a memory your AI tools can find again when it matters.

Start free