Do ChatGPT Projects Share Memory? How Project Memory Works

ChatGPT projects can share context, but only through memory, and only when the project’s memory setting allows it. Files and instructions never cross: a file in one project is invisible to another. Memory is the one channel between them. With default memory, a project’s chats use your saved memories and, on most plans, can draw on your other conversations, including other projects, and be drawn on by them. With project-only memory, the project is sealed both ways: no saved memories in, nothing from outside, and nothing leaking out. A shared project is always project-only. So the honest answer to “do my projects share memory?” is: check each project’s setting, because the answer is per project.

Choosing the setting when you create a project, and the rest of the set-up, is in ChatGPT Projects best practices. This post goes one level down: what exactly crosses, and how to check it. It follows OpenAI’s help centre and documentation as they read at the end of September 2026.

Two kinds of memory, one setting

ChatGPT’s memory is two things, and the project setting treats them differently. The help centre describes them like this.

• Saved memories: details you explicitly asked ChatGPT to remember, or that it saved as useful, stored separately from your chats. They belong to your account, not to a project.
• Chat history: with Reference chat history turned on, ChatGPT draws on relevant parts of past conversations. This is not a fixed list; it is a continually updated synthesis, and what it holds can change as ChatGPT decides what is useful.
• The memory summary: a high-level view of that synthesis, so you can review what ChatGPT has taken from your chats.

The project memory setting decides which of these a project’s chats can use. OpenAI’s personalisation guide↗ (opens in a new tab) adds a third, separate thing: custom instructions, your standing preferences for every chat. Project instructions override them inside the project.

What crosses between projects, and what does not

Here is each piece of context, and whether it reaches a chat in another project.

• Project files and sources: never. They are available only to chats in their own project.
• Project instructions: never. They apply only inside the project.
• Saved memories: yes in a default-memory project, no in a project-only one.
• Chat history from other projects and general chats: in a default-memory project, on most plans, yes. Chats can draw on conversations inside and outside the project, unless the other project is set to project-only memory. On Enterprise and Edu, the help centre says default-memory projects stay contained: chats cannot reference conversations outside the project, and outside chats cannot reference them.
• Chat history from this project, seen from elsewhere: in a default-memory project on most plans, other chats can draw on it. In a project-only project, nothing outside can.
• Chats inside the same project: they can draw on each other under either setting.
• Custom instructions: apply in private projects unless the project instructions override them. Shared projects do not use a member’s own custom instructions or memories.

The last line of the project-only definition is the one people miss. It is not only that the project cannot see out; the rest of ChatGPT cannot see in. That is what makes it the setting for client work: what you discuss about one client cannot surface in a chat about another.

Default vs project-only in practice

1. A personal “Kitchen renovation” project on default memory. ChatGPT already knows from your saved memories that you are in the UK and prefer metric. A general chat next week about tiles may pick up budget details from the project’s chats.
2. The same project on project-only memory. Your saved memories are not used, so tell it your location and units in the project instructions. The tile chat outside the project knows nothing about the renovation.
3. Two client projects, both project-only. Neither sees the other, and neither sees your general chats. This is the arrangement OpenAI’s guidance points to for client accounts and sensitive work.
4. Two client projects, both on default memory. On most plans, each can draw on the other through chat history. If you need them apart, this is the setting to change.

There is one cost to project-only memory. The help centre notes that ChatGPT Work is not available in a project set that way, so a project where you want Work to run multi-step tasks has to stay on default memory.

Shared projects are always sealed

When you share a project, the help centre says it is set to project-only memory and cannot be switched back to default memory. Shared projects do not have access to any member’s own context, custom instructions or memories from outside the project. This protects both sides: your saved memories do not colour answers your colleagues see, and their memories do not colour yours.

The flip side is that everything the project should know has to be in the project: its files, its sources and its instructions. Members of a shared project also see its chats, so a chat inside the project is itself shared context for everyone in it. How Claude handles the same question is compared in Claude Projects vs ChatGPT Projects.

How to check and change a project’s memory

1. Open the project, then its settings from the ••• menu.
2. Look under Memory: it says Default or Project-only.
3. To change it, choose the other option and save. For eligible unshared projects you can switch either way without starting a new project. The help centre warns the new setting can take a few hours to take effect.
4. On Business and Enterprise, project-only memory is offered when Memory is enabled for both the workspace and your account. If the option is missing, ask your admin.

How to reset what ChatGPT remembers

There is no per-project memory list to review, so resetting works at the account level and at the chat level.

• Review: open Settings, Personalization, and look at your saved memories and the memory summary.
• Remove one thing: the help centre’s advice is to delete the saved memory and the chat where you first shared the information, and to remove it from anywhere else it appears, such as a project file.
• Remove everything: choose Delete and turn off memory from the menu on the memory page. This does not delete your past chats.
• Stop one chat feeding memory: use a Temporary Chat. It does not use existing memories or create new ones.
• Stop a bad chat influencing a project: delete it. In a project-only project, moving it out also works, because nothing outside can reach in.
• In the ChatGPT desktop app, local memories are separate: the /memories command controls whether a chat can use them or contribute to them, and the setting lives under Settings, Personalization.

What is still unclear in OpenAI’s documentation

Some questions the documentation does not settle. Treat these as unknowns rather than guessing.

• Whether a saved memory created inside a default-memory project is then used in every other default-memory project. Saved memories belong to the account, which suggests yes, but the help centre does not say so directly.
• Which plans can use project-only memory. The help centre describes it as a setting on every project, while the OpenAI Academy guide to projects↗ (opens in a new tab) describes it as available to Business and Enterprise users with Memory switched on. If you do not see the option, that difference may be why.
• How the two memory systems meet. OpenAI’s page on memories↗ (opens in a new tab) says ChatGPT on the web uses ChatGPT memory, while local Codex clients use a separate local memory store, and that ChatGPT Work on the web does not use the local store. A local project in the desktop app can therefore draw on different memories from a web project of the same name.
• What “a few hours” means when you change the setting, and whether chats started in the meantime follow the old setting or the new one.

The safe reading of all four: if context must not cross, use project-only memory and do not rely on the timing; if context must be available, write it into the project rather than trusting memory to carry it.

Keep what must not be forgotten out of memory

Memory is recall, not a record. You cannot see exactly what it kept, and OpenAI’s own guidance for its developer tools is to keep rules that must always apply in written instructions, not memory. The same holds for a team. Decisions, owners and what is left to do belong somewhere everyone can read and correct. fenbs is a small task board that ChatGPT can reach as a custom MCP app where your plan allows it: tasks in lanes To Do, Next Up, In Progress and Completed, a Decisions page for what the owner decided and why, and AI context notes that each connected assistant reads when it calls fenbs_get_context at the start of its work. A board is the same for every assistant and every person, whatever their memory settings say.