How to transfer your ChatGPT memory to Claude

You have spent months telling ChatGPT how you work: your role, your projects, the names it should not misspell, the tone you want. Now Claude is better at the thing you need this quarter, and the first thing it says is "tell me about yourself". The context you built is real, and it is stuck.

This guide is about getting it out. There are three ways, and they are not equivalent: one is a backup, one is a five-minute paste, one is a permanent fix. Pick by how often you expect to switch again.

First, what "ChatGPT memory" actually is

Two different things live under that word, and only one of them is easy to move.

• Saved memories. Short facts ChatGPT decided to keep — "user is vegetarian", "user works in Rust". A few dozen lines. This is what you actually want in Claude.
• Conversation history. Every chat you ever had. Thousands of lines, most of them irrelevant today. Useful as an archive, useless as context.

The official export gives you the second. What you want to carry over is the first. Keep that distinction in mind and the methods below make sense.

Method 1 — the official export (your backup)

OpenAI's documented path: profile icon → Settings → Data Controls → Export → Confirm export. You get an email with a download link; per the same page the link expires after 24 hours, and the export can take up to seven days to arrive. Inside the zip is your chat history as `conversations.json` — or several numbered files for large accounts.

Do this once regardless of what else you choose. It is the only copy of your history that does not depend on OpenAI keeping your account. But do not expect to hand the zip to Claude: it is a tree of message nodes, not a profile of you. Something has to read it and extract the few hundred words that matter.

Method 2 — the five-minute paste

This is what most people should do, and it takes less time than reading this section.

1. In ChatGPT, ask: "Write out everything you remember about me, verbatim, as a list. Include preferences, projects, people, and constraints."
2. Read the list. Delete what is stale or wrong — there will be some. This is the only moment you get to audit what an assistant thinks it knows.
3. In Claude, open Settings → Memory and use Start import: paste the text and click Add to memory. Anthropic's own page describes exactly this flow, and notes the feature is experimental and may not incorporate every line.
4. Alternatively, just paste the list into a new Claude chat and say "remember this".

Two honest caveats from Anthropic's documentation: Claude's import prioritises work-related context and may leave out personal details, and imports are available on Free, Pro, Max and Team plans on web and desktop. So check what landed in Settings → Memory afterwards and add the missing lines by hand.

The limit of this method is not quality, it is time. You have copied a snapshot. From tomorrow, ChatGPT and Claude drift apart again: what you tell one, the other never learns. If you switch once and stay, that is fine. If you use both — or expect to switch again next year — you have just scheduled yourself the same chore.

What Claude offers on its side

Claude's memory is a list of topics you can view, edit and delete in Settings → Memory, separate per project, with incognito chats that leave no trace. It is well designed. It is also, like ChatGPT's, a memory that belongs to the product: it does not export to anything except a block of text you copy yourself, and no other assistant can read it.

That is the structural problem the two methods above work around rather than solve. Every assistant is building a good memory, and each one is a silo.

Method 3 — one memory both assistants read

The durable version is to stop moving memory between assistants and put it somewhere both can reach. A remote memory server that speaks MCP does this: ChatGPT, Claude, Cursor and the rest connect to the same endpoint with an OAuth sign-in, and read and write the same store.

The migration then happens once, and never again — the per-client steps are in the connection guide:

1. Connect the memory to Claude and to ChatGPT — one URL each. No install; ChatGPT needs one extra step to enable writing.
2. Seed it. Take the list from method 2, paste it into either assistant, and ask it to store each point in memory. The assistant calls the memory tools itself; you watch it happen.
3. Done. Tell Claude something new on Tuesday, ChatGPT knows it on Wednesday. Switch to a third assistant next year: connect it, and it inherits everything.

This is what Pryzm Memory is. It also has a plain JSON export and import in its settings, so the memory you build is a file you can take out at any time — which was the whole point of leaving.

Which one is for you

• You are switching once and not looking back. Method 1 for the archive, method 2 for the context. Fifteen minutes total. You do not need a memory product, and we would rather say so than sell you one.
• You use two assistants, or you switch often. Method 3. The paste method will cost you the same fifteen minutes every time, and you will skip it, and your assistants will get dumber than they need to be.
• You work in a team or across devices. Method 3, because a memory that lives on one laptop or in one account cannot follow you.

If you want to compare Pryzm with the self-hosted and open-source alternatives before deciding — some of them beat us on exactly those two points — the comparison is kept honest and dated.