Source: raw/How_to_Transplant_Your_ChatGPT_Brain_into_Claude_simple_updated_version.md (canonical memory process), raw/How_to_Transplant_Your_ChatGPT_Brain_into_Claude_AI_Memory_Transfer.md (original full-export process), raw/How_to_Move_ChatGPT_Projects_to_Claude_Projects_Easy_Transfer_System.md (projects), raw/Claude_COWORK_vs_ChatGPT_WORK_-_Which_should_you_use.md (reverse import, ChatGPT side). The 10-section migration prompt structure comes from ai-research/eliot-prince-ai-recipe-vault-catalog-2026-05-02.md. Auto-caption transcripts; “Co-work” normalized to Cowork, “claude.md” to CLAUDE.md, “Elliot” to Eliot. 2026-10-27: raw/AI_Solves_722_Math_Problems_at_Once_1.8B_Race_to_Build_Virtual_Cell_Elon_Sees_Quadrillion_SpaceX.md (Moonshots, portability habit) and raw/How_A.I._Agents_Are_Going_to_Transform_Our_Lives_Hard_Fork.md (Hard Fork, Muse import prompt).

Creator: Eliot Prince · Platform: YouTube

VideoURLDurationPublished
How to Transplant Your ChatGPT Brain into Claude (simple updated version)https://www.youtube.com/watch?v=96fX_1KXYRw4:082026-03-04
How to Transplant Your ChatGPT Brain into Claude (AI Memory Transfer)https://www.youtube.com/watch?v=WKpCrvwyXB810:222026-02-13
How to Move ChatGPT Projects to Claude Projectshttps://www.youtube.com/watch?v=V21z0aCaYZc6:542026-06-01
Claude Cowork vs ChatGPT Work (reverse import segment)https://www.youtube.com/watch?v=cVHqpZpbG-s50:242026-09-15

Switching from ChatGPT to Claude means moving three different things, and each has its own route: your memory (what ChatGPT has learned about you), your projects (knowledge files plus custom instructions), and optionally your full chat history. Eliot Prince published the process in three passes between February and June 2026. His March update replaced his February export-based method with Claude’s built-in memory import, so that is the canonical memory route here; the February method survives as the deep option for people who want their whole history mined. By September 2026 ChatGPT had added the same move in the other direction.

Key Takeaways

  • Memory: use Claude’s built-in import first (canonical, March 2026). Claude → Settings → Capabilities → “Import memory from other AI providers” → Start import. Claude shows a ready-written prompt. Copy it, run it in ChatGPT, copy the single code block ChatGPT returns, paste it back into Claude, and click Add to memory. Claude merges it with the memories it already has.
  • Why the update replaced the original process. Users of the February method hit three roadblocks, per Eliot: exports too large to handle, exports that never arrived by email, and general friction. The built-in import needs no export at all.
  • What the import prompt asks ChatGPT for. As read on screen: explore all stored memories and learned context, preserve the user’s words verbatim where possible, and group them under instructions, identity, career, projects, preferences, format and output, wrapped in one code block for easy copying.
  • The import is a snapshot, not a transfer. Eliot’s June video calls it “only just a snapshot of your ChatGPT memory.” Projects, their files and their custom instructions do not come across.
  • Check what you are merging. Eliot ran the import on an already tuned Claude account and worried that “outdated data from ChatGPT” would “mess up my Claude.” On a fresh account he calls it “perfect.” Review Claude’s memory after an import on an established account.^[inferred — he voiced the worry but did not show the result]
  • Projects: ChatGPT has no export button, so ask the project to zip itself. Run a prompt inside the ChatGPT project that gathers every knowledge file, copies the custom instructions word for word into a file named CLAUDE.md, and zips everything for download (prompt below). Eliot calls this “a sneaky little hack”: any of these platforms, Claude included, will zip a project’s knowledge base on request.
  • Why name the instructions file CLAUDE.md. In a Claude chat project you paste its contents into the project instructions box and upload the other files. In a Cowork project created from “Use an existing folder”, Claude reads CLAUDE.md in that folder as the project’s instructions automatically, so the unzipped folder is ready to use as is.
  • Convert files to markdown before uploading. Eliot’s tip for PDFs, PowerPoint and Word files: have ChatGPT, Claude or Cowork convert them to .md first. They take less storage and are easier for the model to read than PDFs with images “baked in,” which he says improves result quality and speed.
  • The manual fallback for projects (March video): open each source in the ChatGPT project and download it one by one, then copy the project settings by hand.
  • Full history (the February method, now the deep option). ChatGPT → Settings → Data controls → Export data → Confirm. A download link arrives at your registered email and expires within 24 hours; large accounts can take a while. The zip contains chat.html with your whole chat history plus uploaded files.
  • Chat projects cap the upload; Cowork does not. Eliot says a Claude project accepts “up to 31 megabytes” of chat history (the recipe-vault page says 32 MB^[ambiguous]). His workaround: open chat.html in a browser, select all, and paste it into the project as text content. Cowork took his 65 MB export directly.
  • The “cognitive migration” prompt mines the history into a profile. The recipe-vault version has 10 sections: Identity Profile, Long-Term Vision, Active Projects, Knowledge Domains, Decision-Making Patterns, Behavioral Patterns, Constraints, Output Styles, Open Loops and a Memory Compression Layer. Eliot reports that the Cowork run worked across 23,512 conversations and ended with usage instructions for Claude (“This isn’t a person who needs handholding”).
  • The February pitfall: Cowork wrote a local memory file. When he asked Cowork to add the profile to Claude memory, it built a persistent memory file on his desktop for Cowork instead. His fix was to copy the profile into a normal chat and say “I have this memory extraction from ChatGPT. Can you merge it into my Claude memories?” This predates the September 2026 change in which chat and Cowork share one memory (see Claude Cowork, reported by Matt Wolfe), so it may no longer happen.
  • The reverse route exists too (September 2026). ChatGPT → Settings → Import → “Import from another AI app” detected Eliot’s Claude Cowork setup and pulled in settings, instructions, plugins, skills, project folders, chat sessions and his custom MCP servers, then listed the connectors to reconnect (ChatGPT calls them plugins). Pointing both apps at one shared folder let him run the same “second brain” in either; ChatGPT reads AGENTS.md where Claude reads CLAUDE.md. Details in Claude vs ChatGPT head-to-head.
  • Lock-in is the underlying problem. Eliot: “these providers want to keep you strapped into their platform, so they’re not making it easy.” A portable folder of markdown files, a CLAUDE.md and skills is what made switching cheap for him by September.^[inferred — he presents the shared folder as the reason the two platforms performed alike]

Try It

  1. Move memory (5 minutes). Claude → Settings → Capabilities → Import memory from other AI providers → Start import. Copy Claude’s prompt into ChatGPT, copy the code block back, click Add to memory. Then open Claude’s memory and delete anything stale.
  2. Move one project. In the ChatGPT project, run the zip prompt below. Download and unzip.
  3. Convert heavy files. Ask Claude or Cowork to convert any PDF, PowerPoint or Word file in the folder to markdown before uploading.
  4. Rebuild it in Claude, either way:
    • Chat project: Projects → New project → upload every file except CLAUDE.md → paste the contents of CLAUDE.md into the project instructions → save.
    • Cowork project: New project → Use an existing folder → select the unzipped folder. Check that the instructions loaded by asking what it knows about you or the project.
  5. Only if you want the deep profile: request the ChatGPT data export, give chat.html to Cowork (no upload cap), run a structured profile prompt, then paste the result into a normal chat and ask Claude to merge it into memory.
  6. Keep the folder as your source of truth. If you might move again, or want to run ChatGPT alongside Claude, keep project knowledge and instructions in a local folder rather than only inside one vendor’s project UI.
  7. Keep a portable memory file as you go. After any major set of queries, run one more prompt: “extract all that key information, figure out what the key concepts are, key relationships, key entities, and create a personal ontology … that makes it portable between models.” Save the result in the project folder from step 6. This is Salim Ismail’s habit (Moonshots, raw/AI_Solves_722_Math_Problems_at_Once_1.8B_Race_to_Build_Virtual_Cell_Elon_Sees_Quadrillion_SpaceX.md), so that “I don’t have lock-in.” It also gives you a ready answer if you move again, in either direction.

Why portability now matters (October 2026)

  • Memory is becoming the lock-in. Salim Ismail argues the shift is from social-graph lock-in to a “much more intimate one”, because “your personal memory belongs to an AI company.” In his view the next big contest is about “which one knows you best,” not which model is smartest (Moonshots, same episode). That is opinion, but it is the reason for step 7 above.
  • Competitors now ask for your Claude history. On Hard Fork, one of the NYT reporters got an alert from Meta’s Muse app mid-recording asking to import his Claude data: “Help me get to know you better. Already use another AI? I can give you a prompt to bring it over” (raw/How_A.I._Agents_Are_Going_to_Transform_Our_Lives_Hard_Fork.md). The import flow this article describes for ChatGPT → Claude now runs the other way too, so decide what you want to take with you before you paste an export prompt into any assistant.

Implementation

Tool/Service: Claude (chat projects and Cowork), ChatGPT (projects, data export, import).

Setup: No install. The memory import lives in Claude’s Settings → Capabilities. Cowork’s existing-folder option needs the Claude desktop app for local folders.

Cost: No extra charge is mentioned; every step runs inside an existing ChatGPT and Claude subscription. Eliot’s prompts sit in his AI Recipe Vault, linked from the video descriptions (catalogued here).

Integration notes:

  • Keep CLAUDE.md out of a chat project’s file uploads; it belongs in the instructions field there and in the folder root for Cowork.
  • ChatGPT’s project custom-instruction box is capped at 8,000 characters (Eliot, April 2026 review, where he had to have ChatGPT shorten his Claude instructions to fit). Instructions moving from ChatGPT to Claude are therefore already under that size.
  • The project-export prompt as Eliot read it from his recipe:
I want you to gather all of the knowledge base files in this project.
The custom instructions need to be extracted word for word and placed in
a file named CLAUDE.md, formatted as clean markdown. Do not summarize or
reword them, copy exactly as written. Then take everything, all of the
knowledge base files plus the CLAUDE.md file, and put them in a single
zip folder for me to download.

Open Questions

  • Does the memory import still sit under Settings → Capabilities after the September 2026 merge of chat and Cowork, and does it accept sources other than ChatGPT? Eliot demos only ChatGPT; the label says “other AI providers.”
  • Project upload cap: 31 MB (Eliot, February 2026) or 32 MB (recipe-vault page)? Neither is a first-party figure.
  • Does Cowork still write a separate local memory file now that chat and Cowork share one memory? Not retested.
  • Custom GPTs are not covered. Eliot’s process handles projects and memory only. With custom GPTs scheduled to stop running on 2026-12-11 for affected Enterprise workspaces, their instructions can seed a Claude project or skill (see Custom GPTs retirement); custom actions need a connector or MCP server.
  • Recipe date mismatch. The recipe-vault catalogue dates the “Migrate Your ChatGPT Brain to Claude” recipe 2025-02-13, but the companion video is from 2026-02-13 and mentions Opus 4.6 and Cowork. The catalogue year is probably wrong.