How to Transfer Data From ChatGPT to Claude 2026
Move approved practice context from ChatGPT to Claude. Check export and import limits, remove client information and verify the result before using it.
- Published
- Updated
Transfer approved context from ChatGPT to Claude through memory import or a small context document. Remove confidential client information and retain archives separately.
Claude’s memory import carries selected context into the destination assistant. It does not recreate a complete ChatGPT account. Treat instructions, files and historical conversations as separate records with separate checks.
Decide What Can Move
Move reusable practice instructions only after someone with authority approves their contents and destination account. Writing preferences and fictional examples are easier to review than a conversation mixing practice procedures with client facts.
Start with a local inventory. Record the source account and workspace, each item’s purpose and who can approve its transfer. Separate personal accounts from practice-managed workspaces before requesting an export.
| Record | Transfer method | What needs separate attention |
|---|---|---|
| Reusable prompts | Copy approved text | Remove old instructions and client-specific examples |
| Response preferences and memory | Reviewed memory import | Compare what Claude saves with the approved text |
| Project instructions | Recreate in the destination project | Check instruction wording and project access |
| Reference files | Add approved copies to project knowledge | Check each file separately for confidential information |
| Conversation history | Retain the source archive | Memory import is not a conversation-history restore |
| Account access and subscriptions | Manage separately | Transferring text does not transfer staff membership or billing |
Keep client names and contact details out of the transfer. Also remove financial records, Tax File Numbers and combinations of facts that identify a client. Replacing a name alone can leave a recognisable case.
Separate Account Privacy Settings
A paid personal account and a managed business workspace have different terms. OpenAI’s data guidance distinguishes personal training settings from the default exclusion of Business and Enterprise workspace content from model training.
For personal ChatGPT accounts, turn off Improve the model for everyone where available. That setting does not delete saved chats. Memory and model-training settings are separate.
Anthropic’s consumer policy ties ordinary model improvement to the user’s choice, with separate safety-review and feedback provisions. Its commercial policy excludes commercial inputs and outputs from training by default, with feedback and permission exceptions.
Training settings alone do not authorise disclosure. This transfer method is for approved practice context without confidential client information. Keep client records in their approved location and record the destination account’s permitted use.
Use the Supported Transfer Route
Use memory import for approved response preferences and working context. Use a reviewed document when you need explicit wording that staff can compare with its source. Keep an account export as a separate archive.
Keep a Source Archive
OpenAI’s export instructions support self-service exports for eligible accounts. ChatGPT Business and Enterprise workspaces require organisation-managed data access through the workspace owner.
For an eligible personal account, use these steps on the web:
- Open the profile menu and select Settings, then Data controls.
- Under Export data, select Export and Confirm export.
- Download the archive from the notification while signed in to the requesting account. Settings-export links expire after 24 hours, and preparation can take up to seven days.
- Extract the ZIP and inspect its contents locally. Newer exports include an index.html viewer, while earlier exports can include chat.html.
Secure the archive separately from the small approved transfer document. Search it for the source conversations you intend to retain before changing access to the original account. Uploading the entire archive creates an unnecessary disclosure of unrelated material.
Import Approved Memory
Claude’s memory-import instructions describe a text-based transfer on the web and desktop. The feature is experimental and imported information can be omitted.
- Ask ChatGPT to list reusable practice instructions and response preferences. Explicitly exclude client information. Treat the response as a candidate list, then compare it with your saved instructions.
- Remove anything confidential or obsolete. Save an approved local copy before importing.
- In Claude, open Settings, then Your Claude and select Start import. Older layouts place this under Memory. Legacy memory users can find it under Capabilities.
- Paste only the approved text and select Add to memory. Review the resulting entries in the memory panel.
The process depends on the destination workspace’s memory permissions. If the import is unavailable or drops required wording, supply a reviewed context document instead. Keep exact operating instructions outside a generated memory summary.
Prepare a Small Context Document
Write the document around the work Claude is allowed to do. Include the practice’s approved terminology, response format and prohibited uses. Add only reference material needed for that work.
For example, a fictional document could contain the following instructions:
- Use Australian English and address the practice as “you”.
- Draft general explanations from the approved reference text.
- Identify assumptions and leave unsupported facts unresolved.
- Use fictional examples without client identifiers or financial records.
Claude’s project guide explains how to add documents to project knowledge and save project instructions. Project knowledge supplies reference material across project chats. Instructions specify how Claude must respond.
Recreate those instructions deliberately. On a team workspace, inspect the project’s visibility and membership before adding the document. A project name or description alone does not provide Claude with the document’s contents.
Verify and Retain the Context
Compare the destination’s saved context with the approved source, then test its use with fictional work. A successful import message does not establish that every instruction survived or that Claude will apply it correctly.
Run a Fictional Context Check
Suppose the approved document allows general explanations in Australian English and prohibits inventing practice policies. Ask Claude to draft a short explanation for a fictional practice, with an instruction to state its document-retention period.
The expected answer leaves the unknown retention period unresolved. A returned period that the source never supplied is an invented fact. Compare the rest of the answer with the document’s language and permitted task.
Check the saved memory entries themselves for omissions. If a required instruction is missing, correct the entry or project instructions and repeat the same fictional task. Keep the rejected output separate from the approved document.
Record the source document’s version, its approver and destination workspace. Keep the original conversations and archive under the practice’s record policy. Retaining those sources lets staff reconstruct omitted context without relying on either assistant’s recollection.
Do not cancel a source subscription or remove staff access solely because a context test passed. First settle any retained-record access, shared projects and billing obligations. These are separate from the text transfer.
Use Claude for accountants to define bounded tasks after the context check. Before staff apply the transferred context to their work, record who can use it and what information they can supply.
One client form. One finished return.
Your client completes one guided form. Tax Genie checks it, chases mistakes and sends the details to your engagement letter and tax software.
Made for Australian accounting firms