Copy Your AI's Memory List Into a File You Own

The single habit is small: once a month, open your AI's memory page, copy the list, and paste it into a dated Markdown file on your own disk. Five minutes. The list is the most concentrated profile any vendor holds on you, and the one you most easily forget you can read.

The reason the habit matters has been public since OpenAI shipped memory in February 2024 — "Remembering things you discuss across all chats saves you from having to repeat information and makes future conversations more helpful," the launch post promised¹, and an April 2025 update extended the feature to "ChatGPT Free, Plus, Team, and Enterprise users."¹ In February 2026, TechCrunch reported that ChatGPT had reached "900 million weekly active users," with OpenAI also disclosing "50 million paying subscribers."² That profile now sits beside hundreds of millions of accounts. It can be read on a settings page. It cannot — on three of the four frontier vendors — be downloaded.

The single habit: a one-page profile, in your file system

Once a month, copy your AI's saved-memories list into a plain .md file with the date in its name. Keep one file per account; add a YAML frontmatter line for last-synced so future you can tell the file from the live page. That is the whole discipline — the rest of this post is the per-vendor receipts.

The habit is small because the value is asymmetric. The list took your AI months of conversation to assemble. Copying it takes one paste. Losing it — through an accidental account reset, a switched plan, a deleted thread, or a vendor's quiet UI change — takes one click. And the list is dense in a way the underlying transcripts are not: it is the synthesis the vendor pulled from those transcripts, the working answer to "what does the model think you want." That synthesis is the part you cannot rebuild by re-reading old chats.

The five-minute version: per-vendor sequence

Five vendors, five settings paths, five paste targets. Use this in order; each step ends with the same instruction: select all, copy, paste into your dated Markdown file.

1. ChatGPT (OpenAI). Open chatgpt.com → Settings → Personalization → Memory → Manage memories³. The saved-memories list is the visible column. Select all, copy, paste under an # OpenAI / ChatGPT — saved memories heading in your file. Why not the Data Export? OpenAI's own help article describes the download as "your chat history and other relevant account data"⁴; a verbatim grep against the article finds zero occurrences of the word "memory." The ZIP holds transcripts. The synthesized profile is not in it.
2. Claude (Anthropic). Open claude.ai → Settings → Capabilities → View and edit your memory⁵. Anthropic's help page says it plainly: "You can save this exported memory as a backup or bring it to another AI service by copying and pasting it into a local file on your computer."⁵ Memory import + export shipped on 2026-03-02 for Free, Pro, Max, and Team plans on web and Claude Desktop⁵. Use the in-Settings export.
3. Gemini (Google). Open gemini.google.com → Settings, then the personalization surface that covers "memory of your past Gemini chats" and the separately-listed "Saved info"⁶. Personalization features are limited to consumer Google Accounts; Google's help page notes "These features are not available while signed in to a work, school, or supervised account."⁶ Google Takeout exports your Gemini Apps Activity (the transcripts), not the synthesized memory surface. Copy what the settings page shows you.
4. Perplexity. Open Perplexity → Settings, then the in-product Memory toggle and the AI Data Retention setting⁷. Perplexity's launch post for assistants with memory says: "With memory, Perplexity remembers key details across conversations so future answers are more personalized and efficient," with controls including incognito mode and "you can even opt out of contributing to model improvement via the AI Data Retention settings."⁷ A verbatim grep of the announcement finds zero occurrences of "export." Copy from the in-product surface.
5. Date and version the file. Add today's date to the filename (ai-memory-2026-06-29.md) and a one-line YAML frontmatter (last-synced: 2026-06-29)⁸. The date in the name is the cheapest version-control system that has ever existed; the YAML line makes the file machine-readable to your own tools later.

The whole sequence runs in about five minutes the first time and two the second. Steph Ango, the CEO of Obsidian, named the reason this format choice matters in his 2023 essay File over app: "In the fullness of time, the files you create are more important than the tools you use to create them. Apps are ephemeral, but your files have a chance to last."⁹

The thirty-minute version: a usable, portable profile

When you have thirty minutes instead of five, the same paste becomes a small editing pass that makes the file useful — to you, to a future tool, and to your own audit of what the vendor has decided you are. Treat it like a once-a-quarter chore, not a daily one.

Start by adding a small structure under each vendor heading. Group memories by topic — "projects," "preferences," "identity," "corrections" — so the file reads as a profile rather than a wall of bullets. Anthropic's own description of Claude's memory hints at the natural sections: a memory summary that "captures all its memories in one place for you to view and edit," across response instructions, personal information, projects, technical preferences, and corrections¹⁰. The same categories work for any vendor's list.

Next, do a literal read-through. The point of the saved-memories list is not just portability; it is the only chance you have to see the profile your AI is using to answer you. Simon Willison, writing on his weblog in May 2025 under the title "I really don't like ChatGPT's new memory dossier," framed the unease that read-throughs produce: "As a power user and context purist I am deeply unhappy at all of that stuff being dumped into the model's context without my explicit permission or control."¹¹ The file you just made is the artifact that lets you give or withhold that permission deliberately.

Finally, prune in the vendor, not just in the file. Delete entries you do not want the model to keep on its side too — that requires going back to the same settings page and using the vendor's delete control. Editing the local file is portable; editing the vendor's list is what changes the next answer.

The minimum useful shape for the file is short — eight lines of frontmatter and one heading per vendor, with the categories as ### subheadings underneath:

---
title: AI memory snapshot
last-synced: 2026-06-29
vendors: [chatgpt, claude, gemini, perplexity]
---

# ChatGPT — saved memories
### response instructions
- prefers concise answers, no preamble
### projects
- ...
### identity / preferences
- ...

# Claude — memory profile (from Settings > Capabilities)
...

Two passes later you have a single file that reads in under a minute and tells you what every model thinks it is talking to. The dossier — to borrow Willison's word for it¹¹ — becomes a document you can audit instead of a black box you cannot.

Common mistakes (the ones to avoid)

Five mistakes recur. None of them is a mystery; all of them are easier to make than to recover from.

• Treating the copied file as a security guarantee. It is not. Copying the saved-memories list does not capture the reference-chat-history layer — the inferred memory ChatGPT generates at runtime from past conversations and never displays as a fixed list. OpenAI's Memory FAQ separates the two: there is a "Reference saved memories" toggle and a "Reference chat history" toggle, and "saved memories are part of the context ChatGPT uses to generate a response. Unless you delete them, saved memories are always considered in future responses."¹² The copy is a portable record of what the vendor lets you see. Not a security guarantee — the vendor still has both layers.
• Pasting the file back into a different AI without re-reading what you are handing over. Your saved-memories file is a copy of sensitive content. Treat it like one: re-read it, redact what does not belong in a new vendor's training surface, and remember that the moment you paste it into a new chat, your notes have left your device for that vendor. (Your notes stay on your device; AI features you opt into send the text you choose to your chosen provider.)
• Syncing the file into a third-party note service that re-reads it. A profile is only as private as the storage you put it in. If the destination platform indexes your notes for its own model, the copy you made to escape one vendor's memory just enrolled in another's. Pick storage that is honest about whether it reads your files.
• Letting the file go stale. A six-month-old profile is a curiosity, not a portable record. Re-run the five-minute version on the first of the month; the cadence is the artifact.
• Treating Mem0, OpenMemory, or Perplexity Brain as substitutes. Mem0's open-source repository describes itself as a "Universal memory layer for AI Agents" with 59,000+ GitHub stars as of June 2026¹³ — and it is genuinely useful, but as a developer layer, not an end-user surface for porting ChatGPT memory. The demand signal is real; the consumer product is still you, with a file.

How this works in MNMNOTE

Your saved-memories file is plain Markdown with a date in its name and a one-line YAML frontmatter for last-synced. Anywhere you can open Markdown, you can open this file — the format is what makes it portable to the next tool, and the next tool after that. The editor is downstream of the file; the file is the artifact that lasts.

This post lives in a family. Your transcripts are the what you said axis — covered in Own Your AI Chat History — and the saved-memories file is its diptych on the what the AI thinks of you axis. The persona you give the AI is the upstream pair — A System Prompt Is Just a Persona File — and the persona the vendor builds from you is the one this post tells you to copy. The upstream policy of what never reaches the AI at all is Some Notes Should Never Reach the AI; the read-back discipline is here. The working-memory ceiling that bounds a single conversation is The Context Window Is Working Memory; the persistent-memory hidden corner is this one. Whether any of the above is used to train the model is the separate axis covered in Does Your AI Assistant Train on Your Notes?. The disciplined paste-into-a-dated-file habit family — Keep a Decision Log, Diff Two AI Answers Side by Side, and the YAML frontmatter primer Put Dates on Your Notes — is the same shape applied to a different artifact. The economics that makes "the cheapest memory is the one you already own" literally true lives in Token Cost Is the New Page Count for Your Notes.

Frequently asked questions

How do I export my ChatGPT memory? No native export exists. The Data Export bundle gives you "your chat history and other relevant account data"⁴ — chat transcripts as JSON and account metadata — and a verbatim grep of OpenAI's help article finds zero occurrences of the word "memory." Open Settings → Personalization → Memory → Manage memories³ and copy the saved-memories list into a Markdown file you own. The synthesized profile is observable on the settings page and not in the ZIP.

What does ChatGPT remember about me? Two layers. The saved memories layer is a visible list you can read at Settings → Personalization → Memory → Manage memories³. The reference chat history layer is inferred at runtime from your past conversations and is never displayed as a fixed list; OpenAI's FAQ describes both as user-toggleable, with the dependency that "turning off 'Reference saved memories' in settings also turns off 'Reference chat history.'"¹² You can read one. You can only switch the other off.

How do I move my AI memory between tools? Claude is the only frontier vendor that ships a documented import and export today, at Settings → Capabilities → View and edit your memory for Free, Pro, Max, and Team on web and Claude Desktop⁵. For ChatGPT, Gemini, and Perplexity, the manual copy-into-a-Markdown-file is the honest path. Re-read the file before you paste it into a different AI: your notes leave your device the moment a new vendor receives them.

Can I back up my AI assistant's memory? Yes, manually. Anthropic's own help article instructs the same workflow, naming it directly: "You can save this exported memory as a backup or bring it to another AI service by copying and pasting it into a local file on your computer."⁵ For vendors without an export, the procedure is identical — only the source surface changes.

Where is the ChatGPT memory settings page? Settings → Personalization → Memory → Manage memories³. The same panel houses the Reference saved memories and Reference chat history toggles¹². Vendor UI paths change; if the menu names shift, search the Settings panel for "memory."

Does the ChatGPT data export include my memories? No. The OpenAI help article describes the bundle as "your chat history and other relevant account data"⁴, and a literal grep of the article returns zero occurrences of "memory." The chats are in the ZIP; the saved-memories list is not.

Is Claude's memory transferable? Yes. Settings → Capabilities → View and edit your memory ships on Free, Pro, Max, and Team plans across web and Claude Desktop, with documented import and export⁵. Simon Willison's September 2025 comparison post notes that "Claude's memory feature is implemented as visible tool calls, which means you can see exactly when and how it is accessing previous context," and that the feature is "transparent and can be edited by the user, resolving another of my complaints about the ChatGPT implementation."¹⁴ Claude is the exception that proves the rule for everyone else.

Why not just rely on Perplexity Brain or Mem0? Mem0 is a developer-facing memory layer with 59,000+ GitHub stars¹³, and Perplexity's Brain is a sibling product surface. Both are evidence that the abstraction is hot. Neither is an end-user one-click export of your ChatGPT, Gemini, or Perplexity memory profile. The portable artifact you can carry across vendors today is still a file you wrote.

The list of things your AI knows about you is the only side of the conversation the vendor hands you a window into. Copy what is in the window, date the copy, and put it somewhere the next vendor cannot reach unless you choose. Anthropic is the lone vendor that ships the export today; until the other three catch up, the cross-vendor standard is the file you paste it into.

This post builds on Simon Willison's running coverage of how memory features land¹¹¹⁴ and on Steph Ango's File over app essay⁹ — your notes folder is the same primitive their argument keeps pointing at. The procedure runs in any plain-Markdown editor; a browser-based one with no account and no install is at mnmnote.com.