Copying memory out before you clear it

Most assistants keep some persistent notion of you across conversations, under a name like memory, personalisation, or saved context. It was built incidentally over months, you never wrote any of it, and clearing it is one of the three actions in a switch that destroys something permanently.

So read it before you touch it. It takes fifteen minutes and it is the only chance you get.

Why this is not covered by the export

Whether memory appears in an account export varies by product, and where it does appear it is commonly a thinner artifact than what the product was actually using — a list of stored facts rather than the working state.

Check your export for it specifically rather than assuming either way. If it is absent, the only copy is the one displayed in the account, and the only way to keep it is to read the screen and type it somewhere.

Treat the on-screen version as the authoritative one regardless. It is what the product is using; the exported version, if there is one, is a summary of that.

Read it as an inventory, not as a curiosity

Find the memory or personalisation area of the account settings and read the whole list. It will be longer than you expect and stranger than you expect — a mix of genuinely useful standing facts, a handful of things you mentioned once, and several that are simply wrong.

Copy the whole thing into a plain-text file before you evaluate any of it. Evaluation is a separate step and mixing them means you delete something while deciding, which is the one thing this exercise exists to prevent.

The file goes next to your instruction file, and it is a working document rather than an archive — its job is to be sorted, not to be kept forever.

Sort into three piles

Load-bearing. Facts the assistant used constantly and correctly: what you do, who you work with, the conventions you never restate, the things you would have to explain again from scratch. Usually a surprisingly short list — five to fifteen items.

Incidental. True, recorded, and irrelevant. A city you asked about once, a one-off project. These do no harm on the old side and there is no reason to carry them.

Wrong. Facts that were true and are not, or were never true. This pile is the interesting one, because it explains behaviour you have been working around without knowing why, and it is a good argument for not migrating memory wholesale even if you could.

Restate the load-bearing pile explicitly

The load-bearing items do not go into the new assistant’s memory feature and wait to be learned. They go into your instruction file, as statements.

This is the significant conversion in the whole exercise: an accumulated memory becomes a written instruction. The old assistant knew you preferred metric because you corrected it four times. The new one knows because a line in your instructions says so, and that line survives the next switch too.

Write them as plain declaratives — role, audience, conventions, recurring subjects, the names you use constantly — and keep them in one section so it is obvious later which lines came from this exercise.

Then clear it, if clearing is what you want

Clearing memory on the old account is worth doing in two situations: the account is shared or belonged to an employer, or the stored material is sensitive and the account is going to sit dormant rather than be deleted.

If you are deleting the account outright, clearing memory first is redundant.

Clearing is irreversible. There is no undo, no version history, and no way to ask what was there before. It is the reason the copy comes first, and it is why this is a single sequence rather than two independent tasks.

What moves

WHAT MOVES — memory and personalisation

  · The stored facts, as text you typed out
                    → move, because you copied them by
                      hand. Nothing else does this.

  · The working state the product actually
    used
                    → STAYS BEHIND. What the settings
                      screen shows is a view of it, not
                      a portable form of it.

  · Months of accumulated calibration
                    → STAYS BEHIND. The new assistant
                      starts cold and no file fixes
                      that.

  · Wrong entries you never noticed
                    → STAYS BEHIND, usefully. Do not
                      migrate the errors.

  · Clearing memory
                    → IRREVERSIBLE. No undo, no history,
                      no way to see what was there.

  · Deleting the account
                    → IRREVERSIBLE. Takes memory with it
                      whether or not you read it first.

Expect the new side to feel worse for a fortnight

This is where the cold start is most obvious, and it is worth naming so you read it correctly.

For the first couple of weeks you are not comparing two assistants. You are comparing one that has a year of accumulated context about you against one that has a text file. The gap is real, it is temporary, and it says nothing about the destination — which is why the parallel period should not be used to reopen the decision.

The one thing that shortens it is the load-bearing list, applied deliberately rather than waiting for the new assistant to learn what the old one knew.

What is genuinely lost

The part of the memory that was never displayed. What a settings screen lists is what the product chose to show you as stored facts; the accumulated fit of a long-used account is not reducible to that list, and the difference is exactly what makes a familiar tool feel familiar.

You cannot copy it, no product exports it, and there is no format that would hold it. Fifteen minutes with the list is not a workaround for that. It is the recovery of the small fraction that was written down at all.