You start a fresh chat to plan a client launch, and ChatGPT already knows your name, your tone, and a half-remembered detail from a recipe thread three weeks ago. Then you open a Project and it acts like a different assistant entirely. That is not a bug. ChatGPT runs two separate memory layers: a global layer that follows you across ordinary chats, and a project-only layer that walls itself off inside a single Project. Knowing which one is remembering, and which one is leaking, is the whole game.
Most explainers blur these into one feature called "memory." They are not the same system. They have different defaults, and they answer to different toggles. Get the mapping wrong and you either spill personal context into a work Project or wonder why a Project forgot something you told it five minutes ago in the main app.
Global memory is two things wearing one name
ChatGPT's global memory is actually a pair of mechanisms. "Saved memories" are explicit facts you ask it to keep: your name, that you are vegetarian, that you write in British English. "Reference chat history" is the implicit one, where ChatGPT pulls useful context from your past conversations to shape new ones without you asking. Saved memories persist until you delete them. Chat-history insights drift and update over time as the model decides what is still relevant.
Both layers operate across all of your normal, non-Project chats. Open a new conversation in the main app and that accumulated context is in play by default. You control all of it under Settings, then Personalization, then Memory, where you can toggle saved memories and chat history independently or switch to a Temporary Chat that neither reads nor writes memory.
The two halves behave differently, and the difference matters. Saved memories work like custom instructions that ChatGPT maintains for you: the model writes, merges, and prunes them automatically, and you can open the list to read or delete any entry. Chat-history reference has no such list. It is a softer, model-side judgement about what from your past conversations is worth recalling, which is why it can feel uncannily good one day and forgetful the next. OpenAI also notes that on the Free tier, responses draw from a reduced set of past chats rather than the fuller history available to paid plans.
Why this layer bleeds by design
The whole purpose of global memory is to follow you. That is a feature when you want one assistant that knows your defaults, and a liability when you do not. A detail from a personal thread can surface mid-way through a work conversation, because as far as global memory is concerned there are no walls between your chats. People who treat ChatGPT as a single brain love this. People running separate contexts in one account get caught by it.
Projects group your chats; project-only memory isolates them
A Project is a container: a set of chats plus shared files and custom instructions, all scoped to one piece of work. As of June 2026, Projects are available on every plan tier, including Free, after OpenAI opened them up in early September 2025. The same rollout raised file limits to 5 per project on Free, 25 on Plus, and 40 on Pro, Business, and Enterprise.
On its own, a Project is just organisation. The feature that changes how memory behaves is project-only memory, a separate control. Switch it on and ChatGPT can draw on the other conversations inside that Project for context, but it will not use your saved memories from outside the Project, and it will not carry anything from the Project out into your future chats. That two-way wall is the point.
Project-only memory is a wall, not a folder: nothing global gets in, and nothing from the Project gets out.
The catch most coverage skips: you set it at creation, not after
Project-only memory is a toggle you enable when you create the Project. It does not appear as a switch you can flip on an existing Project later, so if you started a Project without it, you isolate by rebuilding rather than retrofitting. It also depends on global memory being on in the first place: Personal Memory has to be enabled in your settings for the project-only option to exist, and on Business, Enterprise, and Edu plans the workspace-level memory has to be enabled by an admin too.
Here is what most coverage will not tell you. The design intent is full isolation, but the wall is not bulletproof. The writer Stephen Smith documented a case where a user handling multiple legal matters had memory on its default setting and watched ChatGPT open a draft by referencing a different client he had never pasted, uploaded, or mentioned in that thread. The lesson is not that project-only memory fails. It is that "memory" labels several retrieval paths stacked together, and flipping one toggle does not silence the rest. Treat documented isolation as the behaviour to expect, and keep your global saved memories clean so any rare crossover has nothing embarrassing to surface.
Setting it up correctly
Because the toggle is creation-time only, the order of operations matters. Confirm Personal Memory is on in Settings before you start, then create a new Project and switch memory from Default to Project-only in the creation dialog. On Business, Enterprise, and Edu workspaces the admin-controlled workspace memory has to be enabled too, or the option will not appear. OpenAI shipped these project-only memory controls as part of the September 2025 rollout that opened Projects to Free users, at no extra cost.
- Turn Personal Memory on first; the project-only toggle depends on it existing.
- Set project-only memory at creation. You cannot add it to an existing Project.
- On Business, Enterprise, and Edu, make sure an admin has enabled workspace memory.
- Keep your global saved-memory list tidy to minimise any rare crossover into a Project.
How to tell which memory is active right now
The fastest signal is where you are. In an ordinary chat in the main app, global memory is in play and your saved facts and chat-history insights are available. Inside a Project, check how that Project was created. A Project built without project-only memory can still pull from your global context. A Project built with the toggle on cannot, and ChatGPT will anchor only to that Project's own chats and files.
Two quick checks settle most confusion. First, when ChatGPT references something you never told it inside the current Project, that context came from global memory, which tells you the Project is not isolated. Second, ask ChatGPT directly to show what it remembers about you, then compare that against what surfaces inside a given Project. If the isolated Project ignores your global list, the wall is doing its job.
Keep one Temporary Chat in your habits. It reads and writes no memory at all, so it is the clean room for anything you do not want stored in either layer.
Global memory versus project-only memory, side by side
Here is the mapping in one view. "Global memory" means the saved-memories plus chat-history layer running in your ordinary chats. "Project-only memory" means a Project created with the isolation toggle enabled.
| Dimension | Global memory | Project-only memory |
|---|---|---|
| What it remembers | Explicit saved facts plus implicit insights from past chats | Only the conversations and files inside that one Project |
| Scope | All your normal, non-Project chats | A single Project, walled off from everything else |
| Isolation | None: context flows freely across chats | Two-way: global stays out, Project stays in |
| How you turn it on | On by default once Memory is enabled in Settings | A toggle set only at Project creation, not retrofittable |
| Persistence | Saved memories until deleted; chat-history insights drift | Lives and dies with the Project's own chats |
| Best use | Personal continuity: tone, preferences, ongoing context | Sensitive or separate work you want quarantined |
Which one should you actually use
Use global memory when you want one consistent assistant that knows your defaults everywhere: your writing voice, your stack, your dietary rules, the running context of your life. Use project-only memory when separation matters more than continuity. Consultants juggling competing clients, anyone mixing work and personal use, and people handling anything they would not want surfacing in an unrelated chat all benefit from the wall.
Running two clients in one ChatGPT account? Give each its own Project with project-only memory on at creation, so client A's context can never surface inside client B's chats.
The combination people miss: you can run global memory for your personal life and project-only memory for the work that needs quarantine, at the same time, in the same account. They are not mutually exclusive. The mistake is assuming a plain Project gives you isolation. It does not. Only the project-only toggle does.
Three concrete patterns
A freelance designer keeps personal preferences in global memory so every casual chat knows their style, then spins up one project-only Project per client so brand A's brief never colours brand B's work. A job seeker drafts applications in a project-only Project so salary numbers and current-employer details stay out of unrelated chats. A researcher running two sensitive studies gives each its own isolated Project so a finding from one never leaks into the other's analysis. In all three, global continuity and project isolation coexist.
The honest limit: neither memory is yours to carry
Both layers solve a within-ChatGPT problem. Global memory bleeds helpfully across your chats. Project-only memory walls a Project off. But the moment you open Claude or Gemini to do the same work, none of it follows. Your saved facts, your Project context, your accumulated history all stay locked inside ChatGPT. You are renting memory inside one vendor's product, not owning a layer you can point at any model. And the choice ChatGPT hands you is binary: isolate a Project and lose continuity, or keep continuity and lose isolation. There is no setting that gives you portable memory you control across tools, because that was never the product.
Where an external memory layer fits
This is the gap MemX is built for. Instead of choosing between global continuity and project isolation inside a single tool, MemX is one memory layer you own that spans both modes and works across ChatGPT, Claude, and Gemini. You keep your context in one place, decide what each assistant sees, and carry it between tools instead of rebuilding it each time you switch. It is private by architecture: per-user isolation, encryption at rest, CMEK, and your data is not used for training. The trade ChatGPT forces, isolated or portable but never both, is the trade MemX removes.
Frequently asked questions
01Does ChatGPT have two kinds of memory?
Effectively yes. Global memory covers saved facts and insights from past chats across your normal conversations. Project-only memory is a separate, isolated layer that lives inside one Project and does not share context with the rest of your account.
02Do ChatGPT Projects share memory with regular chats?
A plain Project can still draw on your global memory. A Project created with the project-only memory toggle does not: it ignores your outside saved memories and never carries its own context into chats beyond the Project.
03Can I turn on project-only memory for an existing Project?
No. The project-only memory option appears only when you create a new Project. To isolate work that already lives in a normal Project, recreate it as a new Project with the toggle enabled at setup, then move the chats over.
04Is project-only memory available on the free plan?
Projects are available on every tier, including Free, since September 2025. Project-only memory requires Personal Memory to be enabled, and on Business, Enterprise, and Edu plans the workspace memory must be on as well.
05Can I move my ChatGPT memory to Claude or Gemini?
Not natively. Both ChatGPT memory layers stay inside ChatGPT. To carry context across assistants, you need an external memory layer like MemX that you own and connect to each tool yourself.
The takeaway
ChatGPT does not have one memory. It has a global layer that bleeds across your chats and a project-only layer that walls a single Project off, set at creation and dependent on global memory being on. Pick global for personal continuity, project-only for quarantined work, and run both at once when you need to. Just remember the ceiling: every byte of it stays inside ChatGPT. If you want memory that follows you across ChatGPT, Claude, and Gemini, you need a layer you own, not one you rent.
