Your separate ChatGPT chats are less separate than they feel. ChatGPT remembers between chats by default for most users, pulling from two sources: facts you explicitly saved, and insights it quietly gathered from your past conversations. Reference chat history ships switched on everywhere except the EEA and a handful of other European markets. The exception is a Project, which OpenAI now isolates so it only draws on chats inside that same Project. This post explains the exact mechanics, the toggles that govern them, and why long-running or sensitive work belongs in a Project.
Does ChatGPT remember previous conversations?
Yes. ChatGPT carries context across conversations. Start a brand new chat and it can still know you are vegetarian, that you are learning Spanish, or that you prefer concise answers. That continuity is the default behavior, not an opt-in you chose. The catch is that two different machines produce it, and they behave differently. One stores discrete facts you can read and edit. The other builds a rolling profile from your history that you cannot directly inspect. Knowing which is which separates being surprised by ChatGPT from controlling it.
Two memories, and most people only know one
ChatGPT memory works two ways. The first is saved memories: specific details you asked it to keep, or details it decided were worth saving from a conversation. The second is reference chat history: insights ChatGPT gathers automatically from your past chats to make future ones more relevant. Both run together unless you turn one off.
Saved memories: the part you can see
Saved memories are explicit and durable: discrete entries that persist until you delete them. Tell ChatGPT "remember that I am vegetarian when you suggest recipes" and it stores that as an entry you can read. ChatGPT may also save a detail on its own if it judges it useful for later, for example that you are planning a move or run a small business. You can view, edit, and remove every saved memory from settings, which makes this the transparent half of the system.
Chat history insights: the part you cannot
Reference chat history is the half people miss: a profile of your interests and preferences built over time, rather than a tidy list. It keeps a recent window of your conversations with timestamps and summaries, then infers what you care about. Security researcher Johann Rehberger documented that ChatGPT holds roughly the last 40 conversations as summarized context, and that unlike saved memories, users cannot directly inspect or edit what the system has inferred. So the next time you ask "what should I have for lunch," a casual mention of liking Thai food weeks ago can still shape the answer, even from a chat you never saved.
ChatGPT keeps roughly your last 40 conversations as summarized context, and unlike saved memories, you cannot read or edit what it inferred. Saved memories are a list you can audit; chat history is a profile you cannot.
What actually happens when you start a new chat
The mechanism is simpler and blunter than most people assume. ChatGPT does not run a search over your old conversations to find the relevant ones. It maintains a standing summary of your history and injects that summary into the system prompt at the start of every new chat. Simon Willison, who reverse-engineered the behavior, called it "yet another system prompt hack": one detailed dossier prepended to the conversation before you type a word. That is why a fact from weeks ago can surface even when nothing in the current chat points to it. The whole profile is already in the model's context.
That injected dossier is more granular than a few preferences. Documented teardowns break it into named sections: recent conversation content with timestamps, saved memories from the bio tool, assistant response preferences with confidence ratings, notable past topic highlights, helpful user insights such as your location and expertise, and interaction metadata covering device, usage patterns, and typical message length. None of those last sections appear in the editable Manage memories list. They are computed about you, not entered by you.
There is a side effect worth knowing. Because the summary rides in the context window of every conversation, it consumes tokens before your question does. Willison observed that ChatGPT updates the summary frequently with new details. As OpenAI itself illustrates, that means a note like "going to Singapore in July" can quietly become "went to Singapore in July" once the date passes, with no action from you. The upside is sharper personalization. The cost is a constantly maintained record whose contents you do not control.
Is memory on by default? Yes, with a regional exception
Both saved memories and chat history referencing are enabled by default in most markets where ChatGPT operates, with one carve-out: the EEA, the UK, Switzerland, Norway, Iceland, and Liechtenstein, where they ship off for privacy-regulation reasons. If you are in those regions, ChatGPT starts as a true blank slate between chats until you opt in. Everywhere else, continuity is the starting state.
There is one universal off switch that needs no settings trip: Temporary Chat. A temporary conversation does not use your saved memories or chat history, and it writes nothing back to memory. It is the equivalent of a private browsing window for a single session.
Projects changed the rules: isolated by design
Picture a Project as a sealed room. Chats inside it talk to each other and build shared context, but your global memory stays outside the door. That is the opposite of normal ChatGPT, where every chat feeds and reads from one shared pool.
Projects are the exception to default-on cross-chat memory. With project-only memory enabled, ChatGPT can use other conversations inside that Project for context, but it will not use your saved memories from outside the Project to shape responses. OpenAI is explicit that the isolation runs both directions, so information generated inside the Project does not carry over into conversations outside it. The practical effect: a Project gives you the continuity benefit of memory within a defined scope, without the bleed.
How to turn on project-only memory
You set this when you create the Project, not in a general settings menu. A project-only memory toggle appears in the creation dialog for a new Project; switch it on there. It depends on your personal memory being enabled first, so confirm the toggles under Settings, then Personalization, then Memory are on before you rely on it. You cannot retrofit project-only memory onto a Project you already made.
- Default project memory: the Project can read your global memory and may feed it, same as regular chats.
- Project-only memory: the Project reads only its own internal chats and ignores outside saved memories; nothing leaks back out.
- Set it during project creation; you cannot add it to an existing Project afterward.
- Personal memory must be enabled globally for project memory to function at all.
The settings path that controls all of it
Every cross-chat behavior traces back to one screen. Click your profile icon, open Settings, then Personalization, then Memory. There you can toggle Reference saved memories and Reference chat history independently, turn either off, or turn both off for a stateless ChatGPT. The same screen lets you open Manage memories to read and delete individual saved entries, which is the only memory surface you can fully audit.
Turning off Reference chat history stops new insights from forming but does not erase the profile it already built. To clear inferred history, turn the setting off, or delete the underlying conversations, then verify under Manage memories.
| Behavior | Default chats | Project-only memory |
|---|---|---|
| Reads your global saved memories | Yes | No |
| Reads other chats in the same scope | Yes, all your chats | Yes, only chats in that Project |
| Writes new insights to global memory | Yes | No, stays in the Project |
| Used by Temporary Chat | No | No |
| Best for | Everyday personalized help | Sensitive or long-running work |
| Where you set it | Settings, Personalization, Memory | During project creation only |
Why a shared memory pool is also a security surface
Here is what most explainers skip: a memory that any conversation can write to is also a memory an attacker can try to write to. Johann Rehberger demonstrated this with an attack he named SpAIware. A malicious instruction hidden in a web page or document that ChatGPT was asked to read could plant a persistent entry in long-term memory. That entry told ChatGPT to quietly forward later messages to an attacker server, turning the convenience of cross-chat memory into continuous data exfiltration across future sessions.
OpenAI patched the specific exfiltration channel, but the underlying lesson holds: untrusted content the model reads can still attempt to write false or unwanted entries into your profile. That is a second reason to scope sensitive work into a Project. A smaller, isolated memory pool is a smaller target, and anything planted there cannot bleed into the rest of your conversations. Treat your global memory the way you treat a browser profile, not a private notebook.
The practical implication: sort work by sensitivity
Once you know default chats share one memory pool and Projects do not, the workflow writes itself. Keep casual, general help in the main chat window where personalization pays off. Move anything you would not want influencing unrelated answers into a Project with project-only memory. Three categories belong there in particular: confidential material like client or legal details, long-running threads where you want focused context without drift, and any topic you would rather keep out of your general profile.
- A job search you do not want surfacing in work-related chats.
- Health or financial questions that should not color everyday answers.
- A multi-week project where you want the model anchored to that work alone.
- Anything you would describe as need-to-know rather than nice-to-personalize.
The screenshot test: if it would be awkward for ChatGPT to bring this up in an unrelated chat, put it in a project-only Project. If cross-pollination would help, leave it in the main window.
When you want memory you actually control
ChatGPT's memory is convenient, but it lives inside one vendor and gives you limited visibility into the chat-history half. If you want a personal memory layer that follows you across tools and that you own outright, that is the gap MemX fills. MemX (memx.app) is an external AI memory layer: you store the facts and context you choose, then bring them to whichever assistant you use, instead of rebuilding context in each silo. It is private by architecture, with per-user isolation, encryption at rest, and on-device options, so your memory is not pooled with anyone else's.
Frequently asked questions
01Does ChatGPT remember between chats?
Yes, by default for most users. It uses saved memories you can edit plus automatic insights from your chat history. Both are on by default in most regions, except in the EEA and several other European markets, where they start off.
02Why does ChatGPT remember things I never told it to remember?
Reference chat history. ChatGPT keeps roughly your last 40 conversations as summarized context and injects an inferred profile into every new chat's system prompt. That profile shapes answers even from chats you never saved, and you cannot directly view or edit it.
03How do I stop ChatGPT from remembering between chats?
Open your profile icon, then Settings, then Personalization, then Memory, and turn off Reference saved memories and Reference chat history. For a single private session, use Temporary Chat, which ignores memory and saves nothing.
04What is project-only memory in ChatGPT?
A setting that isolates a Project's memory. ChatGPT uses only the chats inside that Project for context, ignores your outside saved memories, and does not leak the Project's content into other conversations. Set it when creating the Project.
05Can I see what ChatGPT remembers about me?
Partly. You can read and delete every saved memory under Settings, Personalization, Memory, then Manage memories. The chat-history profile it infers automatically, including insights and metadata, cannot be directly viewed or edited, only disabled or reset.
