Guides / One memory for ChatGPT and Claude.

1727 words | Esment

One memory for ChatGPT and Claude.

ChatGPT and Claude keep separate memories. Copy what one knows to the other once, or connect both to one live memory, then test that it works.

╌╌╌╌

ChatGPT and Claude do not share memory. Each keeps its own, inside its own account, and neither can read the other. To share memory between them you have two options: copy what one knows into the other once, or connect both to a third memory that they read and write over MCP.

The first option is a snapshot. It takes a few minutes and starts going stale the next day. The second is live: a fact saved from one assistant can be found from the other, but it takes more setup and depends on your ChatGPT plan. If you are moving from one assistant to the other, you want the first. If you will keep using both, you want the second.

One-time importLive shared memory
What you getA copy of what one assistant knew that dayOne memory that both assistants use
New factsStay in the assistant that learned themAvailable to both, if each one saves and looks
CorrectionsMust be repeated in each assistantMade once
SetupCopy and pasteA memory server and a connector in each assistant
Best forSwitching assistantsUsing both every week

Option 1: copy your memory once

From ChatGPT to Claude

Claude has an import flow for this. Anthropic lists it for every Claude plan, including Free, on the web and in Claude Desktop.

  1. Open the memory section of Claude’s settings and start the import. Anthropic’s page also gives the prompt to use.
  2. Paste that prompt into ChatGPT. It asks ChatGPT to list what it has stored about you in one block.
  3. Read the result and remove anything you do not want to carry over.
  4. Paste the text into Claude and add it to memory.
  5. Review the entries Claude created.

Two limits are worth knowing. Anthropic calls imports experimental and says Claude may not always take them in. And Claude’s memory focuses on work topics, so personal details that are unrelated to work may be dropped.

From Claude to ChatGPT

We could not find an equivalent import tool in OpenAI’s help pages, so this direction is manual. Ask Claude to write out its memories word for word, which Anthropic documents on the same page. Then paste the list into a ChatGPT chat and ask it to remember those points. Afterwards, open ChatGPT’s memory settings and check what it actually kept. For the few facts that never change, custom instructions are a steadier home.

Option 2: connect both to one memory

MCP is an open standard for connecting AI apps to outside tools and data. A memory server built on it gives an assistant tools to save and search. Add the same server to ChatGPT and to Claude, and both are working from one store.

A memory saved in CHATGPT is found later in CLAUDE, because both connect to one storeCHATGPTCLAUDEONE MEMORY SERVERSAVEFINDPrefers replies in British English

Three things decide whether this works for you.

The server must be reachable from the internet

claude.ai and ChatGPT on the web do not run on your computer. Anthropic says plainly that Claude connects from its own cloud, so a server on your laptop or behind a VPN will not connect without a public address. ChatGPT works the same way. A hosted server has that address already. A server on your own machine needs a tunnel.

Your plan must allow it

On the Claude side the answer is simple. Anthropic lists custom connectors on all plans, with Free limited to one, and on Team and Enterprise an owner adds the connector for the organization.

On the ChatGPT side, check before you spend time on setup. OpenAI’s help page on custom MCP servers is the source to trust here, because third-party guides disagree with each other. As of this writing it places full MCP support on Business and Enterprise/Edu, describes read and fetch access for Pro, and does not settle every plan. Access may change, so read the page on the day you set this up.

The difference matters for memory. A read-only connection could let ChatGPT recall what Claude saved, but not save anything itself, which makes sharing one-way. On a work account, your admin may also control whether you can add a server at all.

Connecting does not mean it will be used

An assistant decides for itself whether to call a tool, and often it does not think to. ChatGPT also asks you to confirm write actions by default, so expect some prompts when it saves. Most setups need a short instruction telling the assistant when to search and when to save.

Test it before you rely on it

A short test tells you more than any product page, including this one. Pick a harmless fact neither assistant knows yet, such as “weekly reports go out on Thursday”, and use a new chat for every step.

  1. Before: ask each assistant when your weekly reports go out. Both should say they do not know.
  2. Save: tell Claude the fact and ask it to remember.
  3. Cross over: ask ChatGPT the same question, without mentioning memory.
  4. Ask directly: if it did not know, tell it to check your memory, then ask again.
  5. Update: tell ChatGPT the reports moved to Friday, then ask Claude.
  6. Forget: delete the fact and ask both one last time.

With a one-time import, the fact stays in Claude at step 3 unless you copy it across again, and that is the design. With a live memory, step 3 shows whether recall happens on its own and step 4 shows whether the connection works at all. If step 4 passes and step 3 fails, the connection is fine and the assistant needs better instructions. Step 5 catches a read-only connection. Step 6 shows how many places a fact really lives in.

What to think about before sharing

  • Everything an assistant reads from your memory is processed by that vendor under its own policy, wherever the memory is stored.
  • Custom servers are not vetted for you. Anthropic’s advice is to connect only servers you trust and to review the permissions they request.
  • Copies are harder to delete. OpenAI notes that a memory can also live in the chat where you first shared it, and each pasted copy adds another place to clean up.
  • Keep secrets out of it. Passwords and API keys do not belong in any assistant’s memory.

Where Esment fits

Esment is a memory layer for the second option. It is not an assistant and has no model of its own. It gives Claude, ChatGPT and other MCP assistants one lasting memory, and each memory shows which assistant saved it, so a fact saved from ChatGPT says so when you find it later. It runs as a Mac app or as a web dashboard, and there is no Windows or Android version. Plans are listed here, and the Mac app needs a paid one.

The setup differs by assistant:

  • Claude on the web takes a custom connector. Claude Desktop on a Mac has a Configure button in Esment. Claude receives Esment’s instructions when it connects, so there is nothing to paste.
  • ChatGPT takes a custom MCP connector, and you may need developer mode on. ChatGPT’s menus change often, so follow what your account shows. We have not verified which plans allow this, so check OpenAI’s page above first. ChatGPT does not receive the instructions by itself, so you teach it once with a short note.

Neither web assistant is one-click, and recall is not guaranteed in either, which is why the test above is worth running. Memory is added to every turn automatically only in apps with hooks, such as Claude Code.

To reach your memory, the web assistants use either a temporary tunnel from your Mac or the cloud endpoint. With the tunnel, the Mac app must stay running, the address changes on restart and traffic passes through Cloudflare. In the Mac app with sync off, memories are stored on your Mac and do not reach Esment’s servers. Cloud memory is hosted in the US, and the privacy page sets out what goes where.

If you have history worth keeping, the Mac app can also import a ChatGPT conversations.json or a Claude.ai export and extract facts from it. This needs a local model or an OpenAI key.

When you do not need any of this

  • You are leaving one assistant for good. Use Claude’s import and stop there.
  • You switch only now and then. A short note you paste at the start of a chat does the job, for free, on any plan.
  • Your ChatGPT plan does not allow custom servers. A browser extension may suit you better. MemoryPlugin works that way, and also offers a Custom GPT.
  • You want an open source tool with files you can edit by hand. Basic Memory stores notes as Markdown and works with MCP clients.

Save your first memory →

Common questions

In:

Does importing my ChatGPT memory into Claude keep them in sync?

Out:

No. An import is a one-time copy. Anything either assistant learns afterwards stays with that assistant until you copy it again. Staying in sync needs a memory both can reach, such as an MCP memory server.

In:

Can Claude read ChatGPT’s built-in memory directly?

Out:

No. ChatGPT’s memory belongs to your OpenAI account and Claude has no way to read it. You can ask ChatGPT to write out what it knows and paste that into Claude, or keep shared facts in a separate memory that both connect to.

In:

Can I share memory on the free plans?

Out:

Partly. Anthropic lists memory import and one custom connector on Claude’s Free plan. On the ChatGPT side, OpenAI’s help page on custom MCP servers does not place full support on the free plan, so the manual copy is the realistic route there. Check that page for your plan, because availability changes.

In:

Does a shared memory replace ChatGPT’s and Claude’s own memory?

Out:

No. Each assistant keeps its own memory alongside it, so the two can disagree. If an old fact keeps coming back after you corrected it, look in the assistant’s own memory settings as well as the shared memory.

Written by the team that builds Esment, so we are not a neutral party. Tools change quickly: check each one before you rely on it.

Get the next chapter by email.

One email per new chapter. Nothing else.

0.00