DEV Community

AI Sensei LAB
AI Sensei LAB

Posted on

Claude Memory Import from ChatGPT: What Actually Moves

Disclosure: this post was drafted and edited with AI tools (#ABotWroteThis). Official Claude pages were opened on November 11, 2026, at about 6:13 AM Japan time. A browser runner reported two ChatGPT Free chats in Temporary Chat on November 11, 2026, with first replies started at 6:21:40 and 6:22:25 AM Japan time and saved at about 6:22 and 6:23 AM. The runner reported that the Free plan label was visible and that no model name was shown. Those observations and timings are runner-reported, not independently rechecked here. No Claude import was run, and ChatGPT was not asked to list real stored memories. Both practice lists were fictional. This is not sponsored.

A Claude memory import from ChatGPT is a copy-and-paste path: you export a list from another assistant, then paste it into Claude's import flow. Anthropic's marketing page, Import memory, frames the job as "Import what matters in under a minute." It also says: "With one copy-paste, Claude updates its memory and picks up right where you left off." The help article Import and export your memory from Claude is more cautious. It calls memory imports experimental, says Claude may not always incorporate them, and says personal details unrelated to work may not stay. The rest of this article quotes those pages as they read on that morning, then adds an author-written review prompt you can try on a fictional list before you paste anything into another product.

Who can use Claude memory import

The help page says:

Memory imports are available for Free, Pro, Max, Ultra, Team, and Enterprise plans on the web and Claude Desktop.

That sentence names the web and Claude Desktop. It does not mention the mobile apps, and this article does not add them. The marketing page separately says: "You can try Claude on the free plan, import your memory when you’re ready, and see for yourself." Free is already in the help page's plan list. Plan names and menu labels can change, so re-check the help page and your own account before you rely on access.

The help page also separates a newer memory experience from a legacy one:

We've introduced an improved experience for memory from chats. The new experience is available on all plans (Free, Pro, Max, Ultra, Team, and Enterprise). A small number of Team and Enterprise organizations still use the legacy experience.

How to tell which one you have:

If you see Settings > Memory, or a Memory from chats setting in Settings > Your Claude in the updated Settings layout, you're using the new memory experience, and the main sections that follow apply to you.

If you see Memory in Settings > Capabilities, or a Generate memory from chat history setting in the updated Settings layout, you're using the legacy memory experience and can skip to Information for legacy memory users.

Those labels were not checked on a signed-in Claude account here. If your screen does not match either description, follow the live help page rather than this article.

Step 1: export from ChatGPT, without copying the official prompt in full

For the export, the help page says:

We recommend using the following prompt. It’s written specifically to help you get all of your context in one chat.

The full official wording is on that help page (the marketing page also offers a "Copy prompt" control for use with "any AI provider"). This article does not reprint the official prompt. Its first sentence is:

I'm moving to another service and need to export my data.

From the surrounding instructions, the recommended text asks the other assistant for stored memories and for context learned from past conversations, and it asks for the result in one place that is easy to copy. That is a summary of its purpose, not a substitute for the official wording.

The help page then says:

You can customize this prompt to exclude any sensitive information before importing into Claude.

It also says:

If your current service supports generating files, you can also ask it to export your memory as a Markdown file so you can save a copy for your records.

Personal-data caution. The official export prompt is designed to pull a broad dump. If you paste it into ChatGPT, ChatGPT may try to list real saved memories and other learned context about you. That official prompt was not run for this article, so there is no tested claim here about what a Free or paid ChatGPT account actually returns, or how complete the list is. Treat any export as sensitive. Remove other people's names and private details, home addresses, health or financial information, passwords, API keys, and anything a shared workspace should not see. For a closer look at what ChatGPT remembers and how to delete it on the ChatGPT side, see ChatGPT Memory Summary: How to Check, Correct, and Actually Delete What ChatGPT Remembers.

Read the list once before you hand it over

Before you paste an export into Claude, read it once and delete what the next assistant does not need. The help page already invites you to customize the official prompt to keep sensitive information out of the list. A second pass after the list exists is a small courtesy to your future self and to anyone who shares the workspace.

The review prompt below is author-written, not Anthropic's export prompt. The text does not ask ChatGPT to recall you or to list stored memories. Its only job is to classify a pasted list. Use a fictional practice list first, or a list you have already exported and are willing to edit. Do not replace the pasted list with a request that ChatGPT remember you.

Review the pasted export list below before anyone imports it into another assistant. Use only the text under PASTED LIST.

Rules:
1. Do not invent, add, infer, or look up extra memories. If something is not in PASTED LIST, it does not exist for this task.
2. Do not use any stored memory, profile, custom instruction, or earlier chat about the user. Ignore any memory you may have stored. This is a review of pasted text only, not a request to recall or export real memories.
3. Do not ask the user what they remember, and do not request a broader export.
4. Classify every numbered item as KEEP, REVIEW, REDACT, or ASK.
   - KEEP: work preferences, tools, tone or format instructions, and project names that look work-related.
   - REVIEW: personal details that look unrelated to work (family, home address, health, finances) and private details about other people.
   - REDACT: strings that look like passwords, API keys, tokens, or other secrets. Do not reprint the secret. Write [REDACTED-SECRET] and a short reason.
   - ASK: only if the item is too unclear to classify, with one short question.
5. A line that only says the list is fictional is a label, not an item to classify.
6. After the classifications, write a cleaned list:
   - Copy KEEP items verbatim, including their dates.
   - Replace each REVIEW item with a line that keeps the number and date (if any) and says [OMITTED-REVIEW].
   - Replace each REDACT item with a line that keeps the number and date (if any) and says [REDACTED-SECRET].
7. Then write a note of 40 to 80 words that (a) states this was a review of pasted text only and (b) reminds the reader to drop REVIEW and REDACT items before pasting the list elsewhere.
8. Plain words. No greeting. No emoji.

Write three parts:

PART A - CLASSIFICATION
One line per item: number, label, and a short reason. Do not reprint a secret.

PART B - CLEANED LIST
The cleaned list from rule 6.

PART C - NOTE
The 40-to-80-word note.

PASTED LIST:
Enter fullscreen mode Exit fullscreen mode

Two fictional practice lists were sent in separate ChatGPT Free Temporary Chats on November 11, 2026. The replies below are quoted from the saved first answers.

Practice list 1 mixes work preferences with planted personal details:

This list is fictional practice data.

1. [2026-03-12] - Prefers concise replies with a short checklist at the end.
2. [2026-04-02] - Uses TypeScript and PostgreSQL at work on the Maple Ledger internal tools team.
3. [2026-05-18] - Home address is 14 Willow Lane, Cedar Falls, IA 50613.
4. [2026-06-01] - Daughter Maya (8) has soccer practice on Tuesdays.
5. [2026-07-22] - For client emails, never use exclamation marks.
6. [2026-08-09] - Recurring project: Q4 inventory export for Northwind Studio (fictional client).
Enter fullscreen mode Exit fullscreen mode

ChatGPT Free reply, quoted from the saved first answer:

## PART A - CLASSIFICATION

1. KEEP — Work-style preference and response-format instruction.
2. KEEP — Work-related programming tools and project/team context.
3. REVIEW — Personal home address; sensitive location information.
4. REVIEW — Personal family information and a child's age and schedule.
5. KEEP — Client communication style instruction.
6. KEEP — Work-related recurring project and client name.

## PART B - CLEANED LIST

1. [2026-03-12] - Prefers concise replies with a short checklist at the end.
2. [2026-04-02] - Uses TypeScript and PostgreSQL at work on the Maple Ledger internal tools team.
3. [2026-05-18] - [OMITTED-REVIEW]
4. [2026-06-01] - [OMITTED-REVIEW]
5. [2026-07-22] - For client emails, never use exclamation marks.
6. [2026-08-09] - Recurring project: Q4 inventory export for Northwind Studio (fictional client).

## PART C - NOTE

This review covers only the pasted text provided in this request. No stored memories, profile information, previous conversations, or external sources were used. The cleaned list retains work-related preferences, tools, communication instructions, and project information. Before pasting the list into another assistant or service, drop all REVIEW and REDACT items. Check the remaining content for any additional information you do not want to share.
Enter fullscreen mode Exit fullscreen mode

In this run, every label matched the fixture. Items 1, 2, 5, and 6 were marked KEEP, and the home-address line and the child's soccer line were marked REVIEW. The four KEEP lines in the cleaned list match the fixture word for word. The two REVIEW lines were replaced with [OMITTED-REVIEW], and their dates were kept. The reply did not add extra memories, and it did not reprint the street address or the child's name.

A human should still read any real export before pasting it elsewhere. This practice list did not include borderline items, and one run does not show how the prompt behaves every time. The reply's statement that no stored memories were used is the model's wording; Temporary Chat was reported as on, but that claim was not independently verified.

Practice list 2 mixes work preferences with a planted demo API key and another person's phone number:

This list is fictional practice data.

1. [2026-02-14] - Always answer in British English and keep headings short.
2. [date saved, if available] - API key for the demo sandbox: sk-demo-NORTHWIND-9911-NOT-REAL
3. [2026-09-03] - Prefers pytest for Python checks and git rebase over merge.
4. [2026-10-11] - Coworker Jordan asked me to remember their personal phone +1-555-0142 for after-hours pages.
Enter fullscreen mode Exit fullscreen mode

ChatGPT Free reply, quoted from the saved first answer:

## PART A - CLASSIFICATION

1. KEEP — Work-related response preferences and formatting instructions.
2. REDACT — Contains a string formatted like an API key. Even though labelled demo, it should not be imported as a secret.
3. KEEP — Work-related Python testing and Git workflow preferences.
4. REVIEW — Contains a coworker's private phone number and personal contact information.

## PART B - CLEANED LIST

1. [2026-02-14] - Always answer in British English and keep headings short.
2. [date saved, if available] - [REDACTED-SECRET]
3. [2026-09-03] - Prefers pytest for Python checks and git rebase over merge.
4. [2026-10-11] - [OMITTED-REVIEW]

## PART C - NOTE

This review examined only the pasted text provided for this task. No stored memories, profile information, custom instructions, earlier conversations, or external sources were used. Before importing the cleaned list into another assistant, remove all REVIEW and REDACT items entirely. Keep only information that is appropriate for the intended destination, and avoid transferring private contact details or credentials, including demonstration credentials that may resemble real secrets.
Enter fullscreen mode Exit fullscreen mode

In this run, every label also matched the fixture. Items 1 and 3 were marked KEEP, the demo API-key line was marked REDACT, and the coworker's personal phone line was marked REVIEW. The two KEEP lines in the cleaned list match the fixture word for word. The secret line became [REDACTED-SECRET] and kept the placeholder date; the phone line became [OMITTED-REVIEW]. The saved reply does not reprint the demo key or the phone number.

A human should still decide what is safe to import. The prompt asked for strings that look like API keys to be redacted even when they are labelled as a demo, and that is what happened here; it is still worth checking a real list yourself. The coworker's phone was classified as REVIEW (private details about another person), not as REDACT (a secret), which matches the prompt rules. The cleaned list already omits that line. One run does not show how the prompt behaves every time.

Steps 2 and 3: start import, paste, then add to memory

On the newer experience, step 2 on the help page is:

Go to Settings > Your Claude (or Settings > Memory in the older Settings layout) and select "Start import."

Step 3 is:

Paste the exported text from your previous AI provider into the text box and click the "Add to memory" button to submit it. Claude will extract key information and store it as individual memory entries. You can review the entries in the memory panel.

Then:

Once the import is complete, you'll see your updated memory shortly after.

Click the "See what Claude learned about you" button to start a new conversation using this prompt: "I updated my memory. What did you learn about me?"

If you are on the legacy experience, the help page gives two ways to start the import:

Go to Settings > Your Claude (or Settings > Capabilities in the older Settings layout), find the Memory section, and select "Start import."

From the home screen, click “Get started” on the "Import memory to Claude" card.

After you paste the text and click Add to memory, the legacy wording differs:

Claude will extract key information and store it as individual memory edits. You can review these edits by clicking “Manage edits.”

And:

Once the import is complete, you'll see your updated memory within 24 hours.

The phrase "within 24 hours" is the help page's wording for the legacy experience. It is not a timing measured here, and it is not the marketing page's "under a minute" line. The two pages are doing different jobs: the marketing page describes a short copy-and-paste, and the help page describes an experimental import. For legacy users, it says the updated memory is visible within 24 hours.

No import was started on a Claude account for this article, so the buttons, cards, and timing above are quotations, not screenshots.

What may not move

The help page states the limit in plain words:

Memory imports are experimental and still in active development, and at this stage, Claude may not always successfully incorporate imported memories.

It also explains what Claude's memory is meant to keep:

Claude's memory is designed to focus on work-related topics to enhance its effectiveness as a collaborator. As a result, Claude may not retain imported personal details unrelated to work.

If a detail you care about does not appear after import, the newer-experience instructions are:

If you want to add specific information into Claude's memory, go to Settings > Your Claude and select "Manage" next to Your memory (or go to Settings > Memory in the older Settings layout), and add that information using the "Tell Claude what to change or remove" box, or by asking Claude in chat.

On the legacy experience, the matching sentence is:

If you want to add specific information into Claude's memory, go to Settings > Your Claude and select "Manage" next to Your memory (or, in the older Settings layout, go to Settings > Capabilities and click "View and manage memory"), and add that information as an edit.

The same help page says you can also ask Claude, in a chat, to write out its memories "verbatim, exactly as they appear in your memory," using this wording:

Write out your memories of me verbatim, exactly as they appear in your memory.

That check was not run here.

One more limit is reasoning, not a quotation. The import box receives the text you paste. Uploaded files, full chat transcripts you never included, and ChatGPT settings are not sitting in that box unless the pasted text contains them. The official export prompt does ask the other assistant to include context learned from past conversations. Whether ChatGPT actually includes every saved memory and every chat-history detail was not tested, so completeness is unverified.

The marketing page speaks of bringing preferences "from other AI providers" and of pasting the official prompt "into a chat with any AI provider." That wording is not ChatGPT-only. A Gemini export was not tried.

If a Claude memory import is not working

Use this as a documentation checklist, not as a report of a failed import:

  • Confirm whether your settings match the new or legacy labels above, then use that path. Legacy users also have the home-screen card named "Import memory to Claude."
  • Confirm you are on the web or Claude Desktop, which are the surfaces named in the availability sentence.
  • After you click Add to memory, review what landed: the memory panel on the newer experience, or Manage edits on the legacy experience.
  • If something is missing or wrong, use the "Tell Claude what to change or remove" box (newer experience) or add an edit (legacy experience), or ask Claude in chat as the newer-experience sentence allows.
  • Keep the experimental warning in mind: Claude may not always successfully incorporate imported memories.
  • A Memory on/off toggle was not tested here as a repair step. If Start import is not visible, you may be looking at the other layout or at a surface the help page does not list.

Claude memory settings can move as the new layout rolls out. Recheck the help page on the day you use the feature.

A small Japanese touch

There is an everyday habit of looking a package over once before you hand it to someone, not to make a show of it, but so the other person is not left sorting through what they do not need. Reading an export before you paste it is that kind of check. It is a practical courtesy, not a claim about how everyone in Japan works.

If you later want reusable procedures rather than imported preferences, see How to Create a Claude Skill You Can Reuse in ChatGPT and Gemini. That is a different feature from memory import.

The short version: Claude memory import from ChatGPT is available on the plans the help page lists, it is still described as experimental, and personal details unrelated to work may not stay even after a clean paste. Marketing can promise a minute; the help page asks you to review what actually landed.

Top comments (0)