Transfer ChatGPT memory to Claude: the prompt, and the 10 topics it filters

How do you transfer ChatGPT memory to Claude?

The import works on the Free, Pro, Max and Team plans on the web and Claude Desktop as of October 2026, and the step that decides how much of you arrives comes before the paste, not after it. Five steps, in this order:

1. In Claude, open Settings, Memory and select Start import. If Memory sits under Settings, Capabilities instead, your account is on the legacy experience: same button, slower result.
2. On the same Settings, Memory screen, decide whether to turn on Include sensitive topics in memory. Do it now if you want health, religion or politics to carry over, because Anthropic says Claude does not save those topics retroactively.
3. Copy the prompt the import screen shows, paste it into a new ChatGPT chat with memory switched on, and send it.
4. Copy the code block ChatGPT returns, paste it into Claude's text box, and select Add to memory.
5. Select See what Claude learned about you, read the answer, and fix anything missing under Settings, Memory, Topics.

The fast result belongs to Claude's new memory experience, and it predates the page's September date. Anthropic's import and export page already gave it in its 16 July 2026 version, when the new experience was rolling out as the default for new users, and the 2 September 2026 edit says the new experience "is available on all plans (free, Pro, Max, Team, and Enterprise)." On it, Claude "will extract key information and store it as individual memory entries", and "you'll see your updated memory shortly after." The 24 hour wait survives only in the page's section for legacy users, the small number of Team and Enterprise organizations Anthropic has not moved yet. Guides written between March and May 2026, when the tool launched (MacRumors reported it on 2 March 2026, alongside memory reaching the free plan), still point at Settings, Capabilities and tell you to wait a day.

The import is the easy part. What decides how much of you arrives is the filter on the Claude side, which Anthropic publishes on its memory page:

The six topics behind the switch are health, race, ethnicity, religious beliefs, politics and gender identity, "and other similar areas". The four refusals are government ID numbers, criminal history, financial account numbers and immigration status, and Anthropic says Claude "will let you know when it can't save something for this reason." Anthropic's import page adds one softer filter on top: Claude "may not retain imported personal details unrelated to work", because its memory is built for collaboration. A ChatGPT dump full of family names and hobbies loses more than one full of projects and tools.

What is the prompt to transfer ChatGPT memory to Claude?

The prompt to transfer ChatGPT memory to Claude is the one Anthropic shows on the import screen, and it is the same on the help page. Here it is as published on 2 September 2026, with one dash changed to a full stop:

```

I'm moving to another service and need to export my data. List every memory you have stored about me, as well as any context you've learned about me from past conversations. Output everything in a single code block so I can easily copy it. Format each entry as: [date saved, if available] - memory content. Make sure to cover all of the following. Preserve my words verbatim where possible: Instructions I've given you about how to respond (tone, format, style, 'always do X', 'never do Y'). Personal details: name, location, job, family, interests. Projects, goals, and recurring topics. Tools, languages, and frameworks I use. Preferences and corrections I've made to your behavior. Any other stored context not covered above. Do not summarize, group, or omit any entries. After the code block, confirm whether that is the complete set or if any remain.

```

It covers six categories and ends by asking ChatGPT whether anything remains, so read that last line before you copy. If ChatGPT says entries remain, reply "Continue the list in a second code block" and paste both blocks into Claude together. If you're moving because the saved list keeps filling up, try the fixes in ChatGPT memory full first. Its merge prompt also leaves you a shorter dump to import.

One gap specific to ChatGPT is worth a second message. OpenAI's memory page lists five sources ChatGPT personalizes from (past chats, saved memories, custom instructions, files in Library and connected apps such as Gmail) and says the memory summary "does not necessarily include every detail or source that Memory can reference." Anthropic's prompt asks for memories and context, not for those sources by name. So in the same ChatGPT chat, send this second prompt:

```

Now add a second code block with three parts. First, my custom instructions exactly as saved, word for word. Second, every item in my saved memories list, word for word, one per line. Third, any detail from my Library files or connected apps that you have used to personalize answers. Put [CHECK] at the start of any line that contains a health detail, an ID or account number, religion, politics or someone else's personal information.

```

Then delete the [CHECK] lines you do not want Claude to hold before you paste. Anthropic recommends the same edit ("You can customize this prompt to exclude any sensitive information before importing into Claude"), and it is easier to cut a line from a text box than to find and delete a memory topic later. For a copy on your own disk, Anthropic also suggests asking the old assistant to export the memory as a Markdown file.

Where does each piece of ChatGPT memory go in Claude?

Each piece of ChatGPT memory has a different home in Claude, and only two of the six travel through the import box. The rest are copied by hand, uploaded again or reconnected. Five rows follow OpenAI's own list of memory sources and one adds ChatGPT projects, all checked against both vendors' help centers on 3 October 2026.

Custom instructions are the row people get wrong. The import prompt does ask for "Instructions I've given you about how to respond", but Claude has a separate field built for standing rules: Anthropic's personalization page describes Instructions for Claude as "account-wide settings" applied to all of your conversations. OpenAI's custom instructions FAQ caps ChatGPT's at 1,500 characters on Free and Go and 5,000 on Plus, Pro, Business, Enterprise and Education, so you are moving a paragraph or a page. Paste it whole.

Projects work the same way one level down. Each Claude Project has its own instructions and, per Anthropic, "its own separate memory space and dedicated project summary", so a ChatGPT project's rules go into project instructions and its history stays behind. Free accounts can create five projects. Your old chats themselves are a separate job with a separate answer; how to switch from ChatGPT to Claude covers the export zip and which apps can read it.

What does not carry over when you move ChatGPT memory to Claude?

Five things do not carry over from ChatGPT memory to Claude, even when the import succeeds. In the order you are likely to notice them:

• Whatever ChatGPT leaves out of its own answer. The dump is ChatGPT describing itself; OpenAI's FAQ says the summary "may omit details that are less relevant", and nothing on the Claude side can recover a line ChatGPT never wrote.
• The 10 categories in the chart above. Six wait for the sensitive topics switch, four are refused outright, and the switch only works going forward.
• Personal details unrelated to work. Anthropic's import page says Claude may not retain them, so your dog's name is less safe than your tech stack.
• Your conversations. The import is memory only; the chats stay in ChatGPT and in its export zip.
• Anything that needs a connection. Library files and Gmail context do not ride along in text form unless you ask for them, and even then you get ChatGPT's notes, not the source.

Anthropic is plain about the reliability of the rest: "Memory imports are experimental and still in active development, and at this stage, Claude may not always successfully incorporate imported memories." Treat the import as a first draft of your memory, not a copy of it.

Two account types change the picture. On Team and Enterprise plans, Anthropic says owners control whether memory is available and "it's off for each member until they turn it on", so a work account may have no Memory screen at all until an owner acts. And Anthropic says memory "isn't available to organizations with HIPAA, public-sector, or custom data retention agreements", where the only route is Instructions for Claude.

How do you check what Claude imported?

You check an import by asking Claude directly, then reading the topic list. The button Claude shows after the import opens a chat with the prompt "I updated my memory. What did you learn about me?", and its answer is the quickest diff against the code block you pasted. For the stored version, open Settings, Memory: everything sits under Topics, and each topic can be edited or deleted. To add something the import dropped, use the box Anthropic labels "Tell Claude what to change or remove", or tell Claude in any chat to remember it.

For a backup of the result, Anthropic's page gives a prompt for that as well: "Write out your memories of me verbatim, exactly as they appear in your memory." Save the answer next to the ChatGPT dump. The two files together are the only portable record of what either assistant believes about you.

Then decide what stays in ChatGPT. Deleting a saved memory there stops it being used, but OpenAI says it "may retain logs of deleted saved memories for up to 30 days for safety and debugging purposes", and turning off Reference chat history schedules what it learned from past chats for deletion within 30 days, while the chats themselves stay until you delete them. How ChatGPT memory works explains why saved memories and chat history need deleting separately.

One pasted instruction block can serve Claude and GPT in a Whizi project

If the reason you are moving memory is that you want Claude without giving up GPT, the import solves half of it. Whizi, which publishes this page, offers a different route: a Whizi project takes up to 32,000 characters of instructions and folds them into every message in that project, whichever model you pick, so the code block from the prompt above can sit there once and be read by Claude Sonnet and GPT alike; Claude Sonnet is on Pro and above. Whizi also imports the ChatGPT export's conversations, which Claude does not (import chats from ChatGPT). Where Whizi loses: an always-on Skill applies instructions to every chat, but one Skill holds 4,000 characters (20,000 across the five a chat can use), so a long memory dump fits whole only in a project's 32,000.