Privacy & Portability
Data portability vs portability of meaning
Short answer
Getting your data back means getting files. Reusing them is another matter: you also need the meaning, which is status, reasons, links between items, the history of replacements and provenance. To find out whether your context is really portable, try to pick up a project with a different AI using only your export.

You decide to switch assistants. You request an export of your conversations, a file arrives, and you figure that is that. Then you open it: hundreds of exchanges in a row, with drafts, dead ends and U-turns. Somewhere in there is the decision that mattered. You can no longer tell which of three versions was the last one.
Data travels well. Meaning does not always. This article pulls the two apart, gives a concrete example of the gap, and suggests a simple test to see where you stand.
Getting your data back is not the same as being able to use it
Data portability is the ability to take away what you produced: conversations, files, notes. That is a good thing, and a necessary one.
But a stack of pages is not a context. To pick up work with another AI, it has to understand what was decided, why, what was dropped, and what still holds. If the export holds only text, it is up to you or the new AI to piece all that together by rereading. In other words, you do the work twice.
When a service hands your data back, ask two questions. Can I open it? And can I get the current state of the project out of it without reading everything? The first is about data. The second is about meaning.
What meaning adds to data
By “meaning” we mean whatever links and qualifies the items of a piece of work. In practice:
- status: decided, hypothesis, abandoned, to be confirmed;
- reasons: why one option was chosen, why another was set aside;
- links: what depends on what, what replaces what;
- history of replacements: what used to be true, and on what date it changed;
- provenance: when the information was recorded, where it came from, whether it was imported.
These are exactly the things you lose when you copy raw conversations. A three-week exchange may contain the decision, but nothing says it is one, or that it has been replaced since.
An example: the same project, two exports
Imagine Inès, who runs a nonprofit’s newsletter. She spent six months with an AI on the sending rhythm, the tone, the sections and the mailing tool. Now she wants to move to another assistant.
First export, raw. A 300-page file (a made-up figure) with the exchanges in order. In March you read: “We keep one send per week.” In May: “Actually, every two weeks, open rates drop on Mondays.” In June: “What if we tried Thursday?” With no labels, nothing says the current rule is “every two weeks, on Thursday, as a trial”. The new AI may well answer “one send per week”, with full confidence.
Second export, structured. A few lines per topic:
Sending rhythm
- Status: trial (since June)
- Current decision: every two weeks, on Thursday
- Replaces: one send per week (March), dropped in May
- Reason: open rates falling on Mondays
- Provenance: decided with the team, noted on 12 June
- Revisit if: open rate does not recover
It is short, and you read it in twenty seconds. The other file takes an evening, with no guarantee.
What a good export contains
Without reinventing the list, here are the main categories: the content itself, dates, provenance, decisions and their reasons, relations between items, the history of replacements, and statuses. According to our documentation, that is what an export of the memory stored in Nawame can contain, along with its essential structure. An export that drops half of these turns back into a stack of pages. Our article on what a real AI memory export should contain goes through each category.
This reasoning applies to any service, not just ours. When you assess an export, look for these categories. If statuses and the history of replacements are missing, you will get data, not meaning.
What does not travel
Some things do not come along, and it is better to say so.
The engine. Whatever ranks, selects and returns context belongs to the service, not to your data. For Nawame, ranking, context selection, heuristics and search logic are not necessarily part of an export. You take your memory with you, not the technology that makes it speak. Another AI, or another service, will therefore find things its own way.
Judgment. An export can say “decided” or “hypothesis”. It cannot say whether that status was right at the time. A status is only what someone wrote down.
What was never written. The decision made out loud, the reasoning that stayed in your head: no export contains them. That is one reason to get into the habit of noting what matters, and only that.
Format matters
The same content can come in three shapes:
- plain text: readable anywhere, but with no structure. You have to guess the statuses and the links;
- Markdown: readable by a person and by an AI, with headings, lists and named fields. A good middle ground for a small volume;
- JSON or another structured format: precise and usable by programs, less pleasant to reread by hand.
None is better in itself. What matters is that the structure lives in the export and not only in the service’s interface. A “status” field that exists only on screen does not travel.
Test your portability in an hour
Rather than trusting a promise, try it.
- Pick a real project, not too big.
- Run the export (or write your brief, if you work without a service).
- Open a fresh conversation with a different AI, with nothing but the export.
- Ask three questions whose answers you know: where do we stand? What did we rule out, and why? Which decision changed recently?
- Write down what the AI got wrong or had to guess.
Each error points to a gap in the export. We walk through the full version, with good practices, in how to move project context from one AI to another. And if you import something into a memory, importing memory without importing hallucinations explains how to do it carefully.
According to Anthropic’s help center, consulted on 30 September 2026, Claude offers to import memory from other assistants and to export its own, a feature described as experimental. The test applies to it like to any other.
Limits of this approach
Meaning cannot be decreed. A structured export is only useful if someone, or something, filled in the statuses and reasons properly beforehand. A “reason” field filled in carelessly is no better than a raw conversation.
Also, the richer the structure, the more upkeep it needs. You have to accept not noting everything. And even done well, a transfer always loses a little: the tone of an exchange, the hesitations, the implicit context.
Finally, we say nothing here about the exact export format of any other provider: that changes, and you should check with them. Full portability does not exist. The aim is a transfer that lets you resume work without rebuilding everything. Details on Nawame’s export are in the documentation, and the trade-off between keeping things at home and delegating is covered in local memory or independent memory.
One thing to try this week
Take a topic you are working on with an AI and write, by hand, a six-line card: status, current decision, what it replaces, reason, provenance, revisit if. If you cannot fill it in, the meaning is not yet in your data. If you can, you already hold the core of what a good export has to carry. And if you are still torn about keeping your context inside the assistant or elsewhere, our piece on why not simply use ChatGPT or Claude memory may help.


