To move selected ChatGPT memory to Claude, review ChatGPT’s saved memories and custom instructions, draft a text summary, remove or correct anything unnecessary, choose the right Claude destination, import the approved text, and verify the resulting entries. This is a manual, user-reviewed transfer of selected context. It does not recreate your ChatGPT conversation archive or provide a direct ChatGPT-to-Claude sync.
Anthropic documents memory imports from another AI provider, but describes the feature as experimental. Treat the import as a draft that needs item-level review, not as proof that every approved detail was retained. A stable preference such as "Use concise headings in research memos" may belong in account memory, while a completed campaign brief should usually stay out of global memory.
The safest operating principle is simple: extract broadly, approve narrowly, and verify at the same level of detail as the information you intended to move.
How to move ChatGPT memory to Claude
Use this sequence: inspect, extract, review, route, import, and verify. You decide what is accurate, current, appropriate to transfer, and where it belongs.
- Inspect: Review ChatGPT’s memory controls and custom instructions separately. Optionally request a full data export as a private archive.
- Extract: Ask ChatGPT, or use Anthropic’s documented source-AI prompt, to produce a structured text summary of useful context.
- Review: Correct stale or inferred claims, remove unnecessary sensitive information, preserve useful provenance, and approve each item.
- Route: Send stable preferences to Claude memory, project-specific material to the relevant Claude project, and temporary directions to the current chat.
- Import: Open Claude’s memory settings, choose Start import, paste the approved text, and select Add to memory.
- Verify: Compare Claude’s resulting entries with your approved list, test a few boundaries, and correct omissions or errors.
There is no documented direct ChatGPT-to-Claude sync or API migration in this workflow. The information that moves is the reviewed text you provide, not the original ChatGPT chat list or its internal memory structure.
Know what you are transferring
ChatGPT saved memories, chat-history references, custom instructions, chat history, and a full account export are different sources or artifacts. They should not be treated as one complete, interchangeable memory file.
- Saved memories and chat-history references: Review the available controls under ChatGPT’s memory guidance. The memory summary may not show every detail ChatGPT can use for personalization, and controls can vary by account, platform, region, and workspace.
- Custom instructions: Review these separately under Settings > Personalization > Custom Instructions where available. OpenAI confirms that custom instructions are included in a data export. See OpenAI’s custom instructions guidance.
- Full data export: Settings > Data controls > Export data is an optional account-data archive, not a prepared Claude memory file. Self-service export availability can differ for personal and managed workspaces. See OpenAI’s export instructions.
A full export can be useful as a private backup or source archive. For Claude’s documented import flow, a curated text summary is the practical input. Neither a summary nor an export deletes ChatGPT memories, chats, files, or other source material.
Extract and clean context before importing
Start a new ChatGPT conversation for a focused draft, or use the extraction prompt Anthropic documents for the source AI. A new conversation can make the request easier to review, but it does not guarantee a complete account of every personalization signal.
Ask for explicit memories, learned context, instructions, preferences, corrections, projects, and dates where available. The following is a suggested editorial prompt, not a vendor requirement:
Draft a structured list of useful context you have about me: saved memories, preferences, response instructions, current role, ongoing projects, and corrections I have made. Include dates or source descriptions when available. Separate explicit user-provided information from context inferred from conversation. Do not decide whether sensitive information should be transferred. I will review and approve every item.
Use the output as a draft. Remove duplicates, completed work, outdated roles, vague claims, unsupported inferences, credentials, account numbers, addresses, confidential company details, and third-party personal information unless transfer is authorized. Apply workplace policy and data-minimization rules before pasting anything into another service.
Keep provenance where possible. Mark whether an item came from a saved memory, custom instruction, user-authored note, or chat-derived summary. Give time-sensitive facts a review date. A simple private ledger can include content, category, source, current date, sensitivity, approval state, and destination scope. These are editorial fields, not a Claude or OpenAI schema.
Use deterministic checks for obvious problems, such as credential-like strings, account numbers, duplicate normalized text, missing categories, and undated current-role claims. Use AI to suggest organization, not to silently decide what sensitive information should be deleted. The account owner or authorized data steward should resolve uncertain, sensitive, or stale items.
Use a defined workflow for each type of context
The extraction process becomes safer when every category has an explicit AI task, human gate, destination, and fallback. The following matrix is a proposed operating model, not a set of vendor modules.
| Trigger | AI job | Validation gate | Action or fallback |
|---|---|---|---|
| Stable response preference | Rewrite the preference as one concise, reusable instruction. | Confirm that it is user-provided, current, and broadly useful. | Approve for Claude account memory. If uncertain, keep it in a private ledger or current chat. |
| Current role or time-sensitive fact | Extract the claim and preserve its source and review date. | Owner confirms accuracy and freshness; reject undated or inferred claims. | Import only after confirmation. Add a future review date or leave it out. |
| Client, campaign, or document context | Summarize the project fact without expanding its scope. | Check authorization, project access, confidentiality, and expiration. | Place it in the relevant Claude project. If project scope is unavailable, use the current chat rather than global memory. |
| One-off task direction | Keep the instruction tied to the immediate task. | Confirm that it should not persist beyond the conversation. | Use the current chat only. Do not import it into account memory. |
Choose the right Claude destination
Claude account memory is not the right home for every useful detail. Claude also offers profile preferences and project-specific instructions or knowledge. Project context is separate from account-wide memory and is not automatically shared across every Claude chat.
- Account memory or profile preferences: Use for stable preferences and broadly useful work habits that remain true across tasks.
- Project instructions or project knowledge: Use for client names, campaign details, editorial briefs, document facts, and other project-specific material. Check project access and sharing before adding it.
- Current chat: Use for temporary directions that should not become persistent context.
A preference such as "Use concise headings" may help across many conversations and suit account memory. A client campaign brief belongs in that project instead. Separating the two reduces stale global context and limits project details to their intended scope.
Claude projects can contain project-specific instructions and knowledge, including uploaded documents and text. See Anthropic’s guidance on creating and managing projects before adding material that should remain project-scoped.
Import the approved text into Claude
Open Claude’s memory settings and choose Start import. Current documentation points to Settings > Your Claude. Some accounts show Settings > Memory, while a legacy interface may show Settings > Capabilities. Paste the approved text into the import box and select Add to memory. See Anthropic’s memory import instructions for the current and legacy paths.
Claude extracts selected information into individual memory entries. The documented flow is available across Free, Pro, Max, Ultra, Team, and Enterprise plans, but interface labels, rollout status, and workspace controls can differ. Anthropic describes the import as experimental and warns that imported information may not always be incorporated. Review the result rather than assuming that every line was preserved.
For organizations designing broader, governed AI workflows rather than moving consumer memory by hand, ConsultEvoAI agent servicesA separate option for designing operational AI systems. It is not a ChatGPT-to-Claude memory migration service.
Verify the result and handle missed items
Use Claude’s import review or memory controls to inspect what it learned. You can also ask Claude to write out what it remembers. Compare each approved source entry with the destination entry, not merely with a confirmation message.
- Check the content: Is the preference or fact present, accurate, and still current?
- Check the scope: Did a global preference land in account memory while project facts stayed in the intended project?
- Test recall: Ask a targeted question about a preference and a project boundary. Check that Claude does not invent details missing from the approved set.
- Resolve exceptions: Correct or add a missing entry through Claude’s memory controls or by asking Claude to update it. Record the outcome as accepted, corrected, rejected, or unresolved.
If an important item is missing, rephrase it as a concise, explicit preference or fact, add it through the available memory controls, and check it again. This is a manual correction path, not an API retry or guaranteed reconciliation process.
- Every approved item has a matching destination entry or a recorded exception.
- A test prompt confirms a stable preference without introducing unapproved facts.
- Project-specific context remains in its project, not global memory by default.
- Stale, incorrect, or sensitive unresolved entries stay out until a person reviews them.
What to do with ChatGPT after the transfer
Generating a summary or requesting an export does not itself delete ChatGPT memories or chats. If you want information removed, manage saved memories and source material separately. Turning memory off is not the same as deleting past chats. OpenAI explains these distinctions in its guidance on deleting and archiving chats and memory management.
Export, memory management, chat retention, and model-training settings are separate controls. Decide which action you want and use the relevant setting. Do not treat the Claude import as a change to ChatGPT data or privacy settings.
Frequently asked questions
Does Claude import my ChatGPT conversations?
No. The documented flow imports text containing memories and learned context into Claude memory. It does not recreate the original ChatGPT conversation archive.
Can I transfer ChatGPT custom instructions?
Yes, as reviewed text. Custom instructions are separate from saved memories, so inspect both sources before preparing the import.
Is Claude memory import available on the free plan?
Anthropic lists Free, Pro, Max, Ultra, Team, and Enterprise as supported plans. Account layout, workspace controls, and feature rollout can still affect what you see.
Will every ChatGPT memory be transferred?
No guarantee is documented. Anthropic describes imports as experimental, so verify the entries you care about individually.
Should I use a full ChatGPT export or a summary?
A full export is an account-data archive and can be retained as a private backup. For Claude’s memory import, use a curated text summary that you have reviewed and approved.
Can I automate this through an API?
The reviewed vendor documentation supports a manual web flow. No official API endpoint, webhook, automated sync, or transactional import operation was verified for this consumer workflow.