ChatGPT's old memory system got a saved fact about you right less than half the time, by OpenAI's own measurement. Claude does not see that memory. Neither does Gemini. Each of the three builds a separate memory of you, and none of them compares notes with the other two. Debug code in ChatGPT, edit your writing in Claude, and let Gemini read your inbox, and you are running three independent relationships with three systems that each learn a different slice of your life and forget that the other two exist.
The cost shows up in small, repeated moments rather than one dramatic failure. You paste your style guide into Claude again because ChatGPT's memory of the project lives in a thread Claude has never read. You describe your tech stack to ChatGPT a third time because Gemini already knows it from a document you wrote months ago, and ChatGPT has no way to see that document. You explain your job title, team, and reporting line to Gemini in a Workspace account that turns out not to carry the same memory as the personal Gemini account you use at home. If you pair ChatGPT for debugging with Claude for code review, or run a personal and a work Google account side by side, this will already be familiar. None of it is an accident. Each company built its memory feature to make its own product more useful and harder to leave, not to make your context portable, and the three systems scope what they remember by different rules.
OpenAI's own numbers: its old ChatGPT memory got a saved fact right 41.5 percent of the time. The June 2026 rebuild, called Dreaming, reaches 82.8 percent.
How ChatGPT, Claude, and Gemini Scope Memory
ChatGPT remembers you account-wide across every chat, Claude keeps memory siloed per project, and Gemini ties memory to one personal Google account that stops working the moment you sign in with a Workspace login. The difference decides whether telling an AI tool something once echoes next week, in the next project, or never again.
ChatGPT: One Memory, Spanning Every Chat
ChatGPT's memory is account-wide by default: anything it learns in one conversation is available in the next, regardless of which chat you open. OpenAI rebuilt the system in June 2026 under the name Dreaming, a background process that reads across a user's chat history and periodically rewrites a readable memory summary, instead of waiting for the model to notice and save individual facts mid-conversation. The result runs on two layers: a list of saved memories you can view, edit, or delete one at a time, and a broader reference to your chat history that the model draws on without a fixed list of facts. That is where the 41.5 to 82.8 percent jump above comes from. OpenAI's own example shows what now updates automatically: "You're going to Singapore in July" becomes "You went to Singapore in July 2026" once the trip has passed. The rebuild reached Plus and Pro subscribers in the US first. OpenAI then extended it to Free and Go users, and to more countries.
Claude: Walled Off by Project, Not by Chat
Claude splits memory into two separate pools instead of one. A project is the Claude feature that groups chats, uploaded files, and instructions around one piece of work. Every project gets its own isolated memory and a dedicated project summary that stays inside that project alone. A second, general memory pool covers chats outside any project. That general pool is shared between Claude's regular chat interface and Claude Cowork when Cowork runs in the cloud; a Cowork session running locally on your own machine does not use memory at all. Either way, the general pool never crosses into project memory. Open a new project for a client and Claude starts from nothing about that client, even after months of discussing them in ordinary chat. Anthropic made memory, including this project-level split, free for all Claude users starting March 2, 2026, alongside an experimental tool for moving memory in and out of Claude. That import tool is manual: you ask another AI assistant, ChatGPT or Gemini included, to list everything it has stored about you in a single block of text, then paste that block into Claude's memory settings yourself. It is a one-time copy, not a live connection, and it only brings memory into Claude. Moving memory back out means repeating the same steps by hand, in the other direction.
Gemini: Account-Wide, Only on Personal Accounts
Gemini's memory lives under a setting Google calls Personal Intelligence, and it works at the account level once enabled. Google's own support documentation states personalization is not available while signed in to a work, school, or supervised account. That excludes a Google Workspace identity, the one most people use for Gemini inside Gmail, Docs, and Calendar on the job, from the same memory layer a personal account gets. Personal Intelligence combines three sources. Saved info is facts and standing instructions you state outright. Chat memory accumulates on its own from past conversations once you turn it on. Connected apps, including Gmail, Drive, Photos, YouTube, and Search history, let Gemini pull in context at answer time once you opt each one in. Google launched Personal Intelligence in January 2026 for Google AI Pro and Ultra subscribers and expanded it to more personal accounts in the US by March.
Two logins, two separate memories, one product name.
In short:
- ChatGPT: one memory, every chat, since June 2026's Dreaming rebuild.
- Claude: a separate memory per project, plus one general pool, free since March 2026.
- Gemini: one personal Google account only, cut off the moment you sign in with a Workspace login.
| Dimension | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Memory scope | Account-wide, spans every chat | Split: isolated per-project memory plus a separate general chat memory | Tied to one Google account; personal and Workspace identities don't merge |
| What it draws from | Editable saved-memories list plus a synthesized reference to past chats | Categorized memory entries Claude writes and updates as you talk | Saved info you state, chat history you opt into, and connected Google apps |
| Moving memory to another AI tool | Settings export is a data-compliance download, not built for reuse elsewhere | Experimental copy-paste import/export, one-time, not a continuous sync | No export of personalization memory to another AI assistant |
| Cost to use it | Included for Plus and Pro at launch, expanding to Free | Free for all users since March 2026 | Requires an eligible personal account; expanded beyond paid tiers in the US |
| Where it stops | Enterprise and Team workspaces handle memory under separate settings | Never crosses between two different projects, or between a project and general chat | Excluded for Workspace, Enterprise, Education, under-18, and several regions |
Why No Vendor Builds the Bridge
This fragmentation is not an oversight. It is the business model. Every fact a model remembers about you raises the cost of switching to a competitor, because leaving means starting that relationship over from a blank context window. A company that let your accumulated preferences walk out the door to a rival chatbot would be handing that rival your switching cost for free, and none of the three vendors here has a revenue reason to do that. The clearest evidence is Claude's own import tool. The only mechanism that moves memory between any of these assistants is a manual prompt-and-paste trick, not an automated sync, and it stops working the moment you have another conversation with the source tool, since nothing updates on either side after that single copy. If any of the three had an incentive to make memory travel freely, this would be a toggle in settings instead of a workaround typed into a chat box. No standards body governs AI memory portability the way email providers settled on common protocols decades ago, and none of OpenAI, Anthropic, or Google has proposed one, because a shared protocol would erase the exact lock-in each company's memory feature is built to create.
The only cross-tool memory feature on the market right now is a copy-paste prompt. That is the clearest sign the fragmentation is a choice, not a limitation.
Three Tells You've Hit the Fragmentation Problem
The problem rarely announces itself. It surfaces as a specific, recurring chore. Three signs mean you are already paying the fragmentation tax.
- You re-paste your style guide or brand voice into a new chat. The tool in front of you has no memory of the one where you explained it.
- You re-describe your stack, conventions, and constraints every time you switch from the assistant you use for architecture to the one you use for a quick fix.
- You've told three AI tools your job title. None of them compared notes, because each one heard it inside a different account, project, or company's product.
What an External Memory Layer Fixes, and What It Can't
An external memory layer does the one thing no vendor has a reason to do: put the same facts and preferences in front of whichever AI tool you open next. MemX works as a memory layer that sits across ChatGPT, Claude, Gemini, and a user's own documents. MemX stores a preference, a style guide, or a standing fact about your work once, and supplies it wherever it is needed, instead of leaving it to be re-typed into each vendor's closed memory. MemX is private by architecture: per-user isolation, encryption at rest, and on-device processing where that applies. It is not end-to-end encrypted and not a zero-knowledge system, and it makes no claim to replace what a single vendor's own memory already does well inside that vendor's product.
It cannot make ChatGPT's Dreaming summary visible inside Claude; that summary lives inside OpenAI's own infrastructure and stays there. It cannot restore Personal Intelligence to a Google Workspace account either, since that restriction is Google's policy choice, not something a third-party layer can route around. It does not replace the native memory inside any one tool: ChatGPT remembering the flow of a single long conversation, or Claude keeping a project's file history straight, still happens where it always did. A fact an AI tool infers on its own mid-conversation, rather than one you deliberately saved, also stays inside that tool rather than flowing anywhere else; an external layer carries forward what you chose to keep, not everything a model happens to guess. What it replaces is the part no vendor will build: one place to keep the facts about you that should not depend on which chatbot happens to be open.
Before switching tools for a new task, check whether the thing you are about to re-explain, a style guide, a stack, a job description, is something you already typed into a different assistant this month. If it is, you are paying the fragmentation tax right now.
01Can I transfer my ChatGPT memory to Claude or Gemini?
Only by hand, using Claude's import tool. Ask ChatGPT or Gemini to list everything it knows about you in a text block, then paste that block into Claude's memory settings. It is a one-time copy, not a live sync, so new memory formed afterward does not carry over.
02Does Claude remember things across all my projects?
No. Each project gets its own isolated memory and summary. General chat memory, shared between Claude's chat and Claude Cowork when Cowork runs in the cloud, stays separate from every project, so a new project starts from zero even if you covered the same topic elsewhere.
03Does Gemini remember me the same way on my work and personal accounts?
No. Personalization and memory are restricted to personal Google accounts. Workspace, Enterprise, Education, and supervised accounts do not get the same layer, so a work Gemini login starts from scratch on anything Personal Intelligence would normally handle.
04Why don't ChatGPT, Claude, and Gemini just share memory with each other?
Memory raises switching costs: the more a chatbot knows about you, the harder it is to leave. Handing that to a competitor would give away the advantage, so none offers a live sync. OpenAI's own data export exists for compliance, not reuse elsewhere.
05What's the easiest way to stop re-explaining yourself to every AI tool?
An external memory layer, such as MemX, stores facts and preferences once, then supplies them to ChatGPT, Claude, Gemini, or your own documents as needed, instead of leaving each one to keep its own separate, closed memory.
