How to migrate your ChatGPT memory (before it rewrites itself again)
In March 2026, Anthropic and Google shipped AI memory import within a few weeks of each other: Claude on March 2, Gemini on March 26. Then in June, OpenAI's Dreaming update turned ChatGPT's memory into something that continuously rewrites itself in the background. So the state of play is strange: your AI memory is more portable than it has ever been, and also more of a moving target than it has ever been.
This is a practical guide to getting your ChatGPT memory out and into a place you can read, correct, and keep. I build one of the destinations I'll mention, so treat that section as the pitch it is. The export steps work regardless of where you land.
Quick answer: ChatGPT does not include your saved memories in its normal data export. To get them, open Settings, go to Personalization, Memory, then Manage memory in a web browser, select all the text, and copy it. From there you can paste it into Claude or Gemini (both added a memory import in March 2026) or into a tool that stores it as individual, sourced facts you can revise. What none of the big assistants preserve is when each memory was learned and why, so capture that copy before the next rewrite.
Why migrate your ChatGPT memory at all?
Three reasons come up over and over.
The first is lock-in. Memory is the thing that makes an assistant feel like it knows you, and it is also the thing keeping you on one vendor. The moment you can move it, switching costs drop.
The second is the rewrite problem. Since Dreaming, ChatGPT no longer keeps a fixed list of things you told it to remember. It maintains a summary that changes on its own. Reporting on the update notes that the settings page shows you some of the result, not a complete ledger. If you want a snapshot of what it believes about you right now, before that summary shifts again, the only way to keep one is to copy it out yourself. I wrote about that audit-trail loss in more detail in what ChatGPT's self-rewriting memory costs you.
The third is simple hygiene. Assistants accumulate stale facts: an old job, a project that ended, a preference you changed your mind about. A migration is a natural moment to read the whole pile and drop what no longer applies.
Step 1: Export your saved memories from ChatGPT
Here is the part that trips people up. ChatGPT has a full data export under Settings, Data Controls, Export data. It emails you a ZIP of your conversation history as JSON. Your saved memories are not in it. The facts ChatGPT has learned about you live in a separate store and do not travel with the standard export (aimemory.pro, memoryx.cc).
To get the memories themselves, do this in a desktop web browser, not the mobile or desktop app:
- Open ChatGPT and go to Settings.
- Open Personalization, then Memory.
- Click Manage memory. This opens the list of what ChatGPT has stored about you.
- Select all of it with Ctrl+A (Cmd+A on a Mac), then copy with Ctrl+C (Cmd+C).
- Paste it somewhere plain, like a text file, so you have a clean copy.
The mobile app doesn't give you a bulk action on your memories, so copying the whole list is really a desktop-browser job. There are Chrome extensions that will export the list to a CSV in one click if you would rather not copy by hand, but a manual copy costs nothing and keeps a third party out of your personal data.
That block of text is your export. Keep it. Everything below is about where it goes.
Step 2: Decide where it should live
You have three realistic destinations, and they are not equivalent.
Another assistant (Claude or Gemini). Both added a memory import in March 2026. Anthropic's is the cleanest: you paste a short prompt it provides into ChatGPT, Gemini, or Copilot, the old assistant packages its memories into a text block, and you paste that back into Claude's memory settings and click to add it (step-by-step walkthrough). Claude's import is free for all users and takes about a minute. Gemini shipped a comparable import around the same time (9to5Google). This is genuinely useful, and if you just want your new assistant to feel less like a stranger, it is enough.
The catch is where it lands. Claude imports your memory as individual memory edits it then manages; Gemini folds it into its own memory. But neither one's documentation describes showing you, per fact, when a detail was learned or from which conversation, or a history of how it changed, and both keep evolving it on their own. You have changed vendors, not gained an audit trail.
A plain file. Save the copied text and you have a durable, private snapshot that no vendor can rewrite. This is underrated. It will not power an assistant, but it is a real backup, and it is the one option where nothing else touches your data.
A memory store you can audit. This is the destination I build, so here is the honest version. In Calmara, you paste the same exported text into a memory import dialog, pick the source, and it extracts the content into individual facts. Every imported fact lands in a pending-review queue rather than going live automatically, so you approve what is true and reject what is stale before it becomes part of your memory. Each fact keeps its provenance and, from then on, a full revision history: what it was, when it changed, and what replaced it. You can retract a fact and keep the history, or hard-delete it entirely. And you can export the whole thing back out in an open format whenever you want to leave.
Step 3: Import into an auditable destination
If you go the Calmara route, the flow after Step 1 is short:
- Open the Memory tab and choose Import memory.
- Pick the source (ChatGPT, Claude, Gemini, or other) and paste your copied text.
- Calmara extracts it into discrete facts, using the AI provider you have configured. Point it at a local model or your own API key and the extraction never touches a shared cloud account.
- Review the queue. Each extracted fact waits for your approval. Approve the accurate ones, reject the rest.
- From then on, every approved fact carries where it came from and a revision history you can open at any time.
The difference from a same-day Claude import is not the 60 seconds it takes. It is what you have afterward: a memory where you can answer "why does it think this?" and "what did it believe last month?" months later, instead of trusting a summary that edits itself while you are not looking. That is the whole reason to prefer an auditable destination, and it is the wedge I keep coming back to: revision history and provenance you can actually prove.
What gets lost in any migration
Be realistic about the ceiling here. None of these methods recover what the source assistant never exposed.
ChatGPT's memory summary is already a compression of your conversations. You are exporting the compression, not the raw material. If a nuance never made it into a saved memory, no import will resurrect it. Migrating memory is also not the same as migrating chat history: your conversation logs export separately as that JSON ZIP. Most tools import only the memory summary, not the conversations; Gemini is the exception and can import full chat history, up to several gigabytes of it (9to5Google).
And timing metadata is gone at the source. Because ChatGPT no longer keeps a dated list of when each memory was added, your export is a flat snapshot. A destination can start tracking history from the moment you import, which is worth a lot going forward, but it cannot reconstruct a timeline the source already threw away. That is the strongest argument for doing this sooner rather than later: the snapshot you take today is the earliest one you will ever be able to keep.
Frequently asked questions
Does exporting my ChatGPT data include my memories?
No. The Settings, Data Controls, Export data ZIP contains your conversation history as JSON, not your saved memories. Memories are stored separately and have to be copied manually from Settings, Personalization, Memory, Manage memory in a web browser.
Can I import ChatGPT memory into Claude?
Yes. Claude added an experimental memory import in March 2026, free for all tiers. It gives you a prompt to paste into ChatGPT, which produces a memory block you paste back into Claude. It imports the memory summary, not your full conversations.
Will a migration keep the dates my memories were created?
No, because ChatGPT stopped exposing per-memory dates after its June 2026 memory rewrite. Your export is a snapshot without a timeline. A destination that records history can track changes from the import forward, but it cannot recover dates the source no longer provides.
Is it safe to paste my memory into another tool?
Your memory can contain personal details, so treat the destination like any place you store personal data: check where it lives, whether it is encrypted, and whether you can delete it. If you use a tool that extracts memory with an AI model, prefer one that runs on your own API key or a local model rather than a shared cloud account, so the text is not pooled with everyone else's.
What is the difference between exporting to Claude and exporting to an auditable store?
Claude and Gemini import your memory into their own stores, and neither one's documentation describes per-fact provenance or version history. An auditable store keeps each fact separately with its source and a revision log, lets you approve facts before they go live, and lets you export them again in an open format. Same 60-second paste, very different thing on the other side.
If your reason for migrating is compliance rather than convenience, the same audit properties are what a review team will ask about. I wrote that up as the six controls your compliance team needs before approving an AI assistant.
Written by Dan Hagen