What does not transfer

People plan a switch around the history and are surprised by the rest. The conversations are the easy part — they export. What actually made the assistant feel like yours is a set of accumulated settings and habits, and almost none of it comes with you.

This is the inventory. It is worth writing your own version of it before you cancel anything, because most items on the list are invisible until they are missing.

Conversation state

The most important item and the one most often misunderstood.

A conversation is not a document with a cursor in it. It is a thread the assistant was following. Exported history gives you the text of what was said. It does not give the new assistant the state of having said it.

You can paste an old conversation into a new assistant, and it will read it the way it would read any pasted document. That is genuinely useful for handing over a long-running piece of work. It is not resumption, and the distinction shows up immediately: the new assistant is summarising a transcript rather than continuing a thread, and its grasp of the thing is exactly as good as the transcript is explicit. Everything that was implicit — established halfway through and never restated — is gone.

No product transfers this. There is no format for it and no reason to expect one.

Memory and personalisation

Most assistants accumulate some persistent notion of you across conversations, under a name like memory, personalisation, or saved context. It is built from months of incidental mentions: your job, the projects you keep returning to, the fact that you prefer metric, that you have a colleague whose name you use constantly.

This does not transfer, and the exported version — if the export includes one at all — is a much thinner artifact than what the product was actually using.

The practical consequence is that the new assistant starts genuinely cold, and the first two weeks feel worse than they are. You are not comparing two assistants; you are comparing one with a year of accumulated context against one with none. That is worth knowing before you conclude the new one is worse.

The one thing you can do about it is deliberate: read the exported memory if you have it, decide which items were actually load-bearing, and state those explicitly in the new assistant’s custom instructions rather than waiting for it to learn them again.

Custom instructions

The standing instructions that shape every response — tone, format, what to assume about you, what never to do.

These almost never transfer, because they are stored per-product and phrased for a particular one. They are also the highest-value thing to move, because they took the longest to get right and they are pure text.

Rewriting them is genuinely worth doing rather than copying verbatim; instructions written as a portable artifact survive the next switch too.

Saved prompts, projects and workspaces

Whatever the product calls the organisational layer — saved prompts, projects, folders, custom assistants configured for a task, workspaces with their own instructions.

None of this has a portable format. Two things are true at once: the content is recoverable if you wrote it down or can find it in the exported history, and the structure is not. You will be rebuilding the organisation by hand, and it is worth taking the opportunity to rebuild only the parts you actually used.

Uploaded files and their context

Files may or may not be in your export. Even when they are, the association between a file and the conversation that analysed it survives only as text.

If you have a working set of documents you re-upload constantly — a style guide, a spec, a reference sheet — treat those as first-class things to keep outside any assistant. That is the version that survives every switch, including the next one.

Integrations and connected accounts

Anything the assistant was wired into: connected storage, calendars, mail, code hosting, third-party tools. All of it is per-product authorisation and has to be redone.

Worth doing in the other order, too: when you stop using an assistant, the authorisations it holds against your other accounts do not expire on their own. Revoking them is part of leaving, and it lives in the other services’ settings, not in the assistant’s.

Shared links and anything anyone else relies on

If you shared conversation links with colleagues, published anything, or embedded a link in a document or ticket, those links stop working when the account closes. Nobody discovers this until a link is clicked.

This is the one item on the list with a deadline that is not yours. If shared links matter, find them and replace them with the exported text before you cancel.

What moves

WHAT MOVES — leaving one assistant for another

  · Conversation text
                    → exports. Readable, searchable,
                      yours.

  · Conversation state — the thread being
    followed
                    → STAYS BEHIND. Pasting a transcript
                      is handover, not resumption.

  · Memory / personalisation
                    → STAYS BEHIND in substance. The new
                      assistant starts cold and the
                      first fortnight misleads you.

  · Custom instructions
                    → STAYS BEHIND, but it is plain text.
                      Rebuild it as a file you own.

  · Saved prompts, projects, workspaces
                    → content recoverable, STRUCTURE
                      STAYS BEHIND. Rebuild only what
                      you used.

  · Integrations and connected accounts
                    → STAYS BEHIND both ways. Reauthorise
                      on the new side; revoke on the old.

  · Shared links you gave other people
                    → break on cancellation. Replace them
                      before, not after.

  · Deleting the account
                    → IRREVERSIBLE. Everything above that
                      you have not already copied out is
                      gone with it.

The useful way to read this list

Not as a warning. As a build order.

Every item that stays behind is either something you rebuild once and keep forever outside any product — instructions, reference documents, the prompts you actually use — or something you accept losing. Sorting the list into those two piles is most of the work of a switch, and it is the part that makes the next one cheap.

The thing you cannot fix by planning is the cold start. The new assistant will know nothing about you for a while, and no amount of preparation makes that feel like the tool you have been using for a year. That is a real cost of switching, it is temporary, and it is worth naming honestly rather than discovering it and blaming the destination.