How to move your AI memory between providers

Step 1 — Export the raw material

Not the memory summary; the conversations. Summaries are the provider's opinion of you and they are always thinner than you expect. The archive is the evidence.

• ChatGPT: Settings → Data controls → Export data. A ZIP arrives by email.
• Claude: Settings → Privacy → Export data. Also read Settings → Memory → Topics and copy the list; it is small and it is the distilled part.
• Gemini: Google Takeout → My Activity → Gemini Apps.

Step 2 — Distil, don't dump

The instinct is to paste everything into the new model. Don't. A year of chats is tens of megabytes, most of it dead — debugging sessions, restaurant recommendations, half-finished drafts. Even where it fits in a context window, filling the window with noise measurably degrades the answers.

What you're mining for is five categories:

• Identity: what you do, where you are, what stage of life you're in.
• Cast: the people who recur, and who they are to you.
• Work in flight: current projects and their state, not everything you've ever touched.
• Standing rules: the corrections you've had to make repeatedly. These are gold and they are the first thing lost in a switch.
• Voice: how you want to be spoken to, and how you write.

Step 3 — Write the brief

300–600 words, written in third person about you, present tense. Third person matters more than it sounds: a model reading "Daniel runs a small software company and prefers direct answers" treats it as fact about the user, where "You are…" reads as a role-play instruction and gets diluted by the system prompt.

Mark confidence honestly. "Probably moving to Lisbon in spring" should not be written as "lives in Lisbon" — that single flattening is how memory systems end up confidently wrong about people.

Step 4 — Verify before you trust it

Paste the brief into the new tool, then ask it three questions you already know the answers to: who are the main people in my life, what am I working on right now, and how do I like to be answered. If it gets those three right, the transfer worked. If it embellishes, cut the embellished lines from the brief — a model will happily build on anything you hand it, true or not.

Doing it once instead of every time

The manual version works and you should know how to do it. The problem is that it decays: the brief is right the week you write it and stale two months later, and nobody rewrites it by hand four times a year.

That's the case for keeping the vault outside any provider — one place that holds the exports, keeps distilling as new material arrives, and hands out a current brief to whichever model you're using this month. See the migration playbook for the long version, and Claude's August 2026 memory update for why editable still isn't portable.