ChatGPT Memory Explained (And When to Turn It Off)

TL;DR: ChatGPT memory is a continually updated synthesis of your past chats, not a stored list, and OpenAI's own documentation admits it can go stale and contradict itself. This guide covers what it stores, where to see it, why it drifts, and the specific cases where turning it off, or scoping it to a project, beats leaving it on.

If ChatGPT has ever brought up an old job, a finished project, or a preference you dropped months ago, you have already met the feature this page is about. ChatGPT memory is what is doing that: a running account of what you have told it, built from every chat on your account, used to shape answers you did not ask it to personalize. It is genuinely useful. It is also, by OpenAI's own admission, not reliable in the way people assume, and most guides to it stop at explaining how to turn it on, without saying when you shouldn't. This one covers how ChatGPT memory actually works, what it stores, why it drifts, and the real reasons to turn it off rather than leave it running by default.

How does ChatGPT memory actually work?

ChatGPT memory works by continually synthesizing your past chats, files, and connected apps into what OpenAI calls a memory summary, then using that summary as context for new answers. OpenAI's own description of it is direct: "When enabled, memory helps ChatGPT automatically remember useful context from your chats, files, and connected apps to personalize your experience, so you don’t have to repeat yourself as often" (help.openai.com/en/articles/8590148-memory-faq, updated 31 August 2026, accessed 4 September 2026).

That single sentence hides an important detail: there are currently two memory systems live at once, and OpenAI's own help page explains why it built the second one. The original system, now called legacy, only stored facts you explicitly told it to remember, and required you to manage them yourself. OpenAI's stated reason for replacing it is worth reading in full: "The previous saved memories system often became stale and relied on users to manually manage updates. Memories could also contradict one another, such as “I’m training for a marathon” and “I sprained my ankle,” which made personalization less accurate." The newer system, OpenAI says, "updates memories automatically with ChatGPT keeping track of the details it determines are most important so it can continue building on the context you’ve already shared."

Both systems still exist, and you can move between them. To go back to the old, explicit-only version: "Go to Settings > Memory > Saved memories to revert to the legacy saved memories system. From saved memories, you can select “Try improved memory” to return to the updated memory experience."

What does ChatGPT memory actually store, and can you see all of it?

No, not completely. The memory summary you can open in settings is a partial view of a larger synthesis, and OpenAI says so directly: "Not necessarily. ChatGPT’s memory is based on a continually updated synthesis of context from your past chats, which may be broader than what can be shown as individual items in a summary." The same page adds that "Some details may not appear in the summary, including when ChatGPT determines they are less relevant or not appropriate to show in this view."

Saved memories specifically get a useful analogy from OpenAI: "The “notepad” of your saved memories is stored separately from your chat history. This means even if you delete a chat, any saved memories from it can still be used in future conversations." Deleted memories are not gone the instant you delete them, either: "We may retain a log of deleted Saved Memories for up to 30 days for safety and debugging purposes."

There is one more place to check, and most people never open it. Under a response, a book icon shows exactly what ChatGPT drew on to shape that specific answer: "You can see what sources were used to personalize a response such as custom instructions, past chats, files, and memories by tapping the book icon below the response. Tapping on a memory in sources will open an explanation on why that memory was used to personalize a response." It has the same honest caveat as everything else here, though: "Sources are designed to make memory easier to understand and control, though they may not show every factor or source that shaped a response."

If you just want a fast answer instead of digging through settings, OpenAI's own suggestion is the simplest one on this page: "If you want to know if ChatGPT has remembered something, just ask in chat."

Audit what memory says about you:

Show me everything you currently believe about me from memory, grouped as:
work, personal, preferences, and anything else. Mark each item as something
I told you directly or something you inferred. Do not soften or summarize
further than that.

Why does memory drift, and why does it contaminate your outputs?

Because it is one synthesis covering every chat on your account, applied automatically, with no built-in boundary between contexts unless you create one. OpenAI's own marathon example is the clearest documented case: a memory that you were training for a marathon and a memory that you sprained your ankle can both be true and both stay in the summary, and the system does not resolve the contradiction for you. That is not a hypothetical; it is the specific failure OpenAI cites as the reason it redesigned the feature.

The same mechanism creates a second, less obvious problem OpenAI does not spell out directly, but which follows from how the system is built: because one memory summary spans every chat on the account, nothing in the interface stops a fact from one context from surfacing in an unrelated one. A detail from a client project can resurface in a different client's chat. A personal preference can resurface in a work chat, or the reverse. OpenAI has not published a statement about this specific scenario, so treat it as a reasoned consequence of the documented mechanism rather than a vendor claim, but the mechanism itself is real: one synthesis, applied everywhere, by default.

OpenAI does flag the closely related privacy version of this problem: "Sensitive information may appear in memory if you share it with ChatGPT. If you don’t want information from a chat to be used to personalize responses, you can turn off memory or use Temporary Chats." And drift does not fully reset even when you try to clear it: "If you turn memory back on later, ChatGPT may create new memories from chats that remain in your chat history, including older chats." Turning memory off and back on is not the clean slate it sounds like.

If ChatGPT seems to be doing the opposite instead, forgetting things you told it rather than remembering too much, that is a different failure mode covered in why ChatGPT forgets what you told it.

When should you actually turn ChatGPT memory off?

Turn it off, or scope it, in any of these cases. None of them are hypothetical; each maps to a mechanism documented above.

1. You use one account across genuinely separate professional contexts — an agency serving several clients, or freelance work plus your own writing. Full memory has no per-client wall by default. A narrower fix than turning memory off entirely is scoping it: OpenAI describes a project-only version of memory where chats inside a project "can reference other conversations in the same project, but they cannot reference conversations outside the project, and chats outside the project cannot reference conversations in it." That is documented specifically in the context of regulated Enterprise and Healthcare workspaces, so treat it as confirmed for that setting rather than promised on every personal plan, but it is the right feature to look for if your account supports it.
2. You are handling sensitive, regulated, or confidential information. OpenAI's own guidance for Enterprise and Healthcare workspaces states that "improved memory is disabled by default" there, and that the improved system "is not covered under your BAA" where a Business Associate Agreement applies. If that describes your account, the default is already off; if you turned it on, know what you turned on.
3. You are testing or evaluating prompts and don't want personalization skewing the output. If you write prompts for a living, or are comparing how a prompt performs, memory quietly changes the baseline every time it injects something you did not put in the prompt itself. Temporary Chat gives you the unpersonalized version on demand.
4. You are about to share your screen, your account, or a chat link. Memory sources are excluded from shared chats, but the answer itself can still reference something you would rather not have surfaced in front of whoever you are showing it to.
5. Memory has drifted stale faster than you are willing to correct it by hand — an old job, a finished project, a preference you changed months ago, still shaping today's answers. Correcting individual memories works, but if it is happening often enough that you are doing it weekly, the setting is costing you more time than it saves.
6. The account is shared or used on a shared device. Nothing about ChatGPT memory assumes more than one person is typing into the same account.

How do you turn ChatGPT memory off, and what does each option actually do?

Start here: "Memory controls are available in Settings > Personalization > Memory. You can choose to enable or disable memory at any time." That single toggle is the account-wide switch. But it is one of four distinct controls, and they do not all do the same thing, which is where most confusion about turning memory off actually comes from.

A few of those rows are worth reading twice. Turning off the account-level toggle stops memory going forward; OpenAI's separate instruction for actually deleting what is stored is a different action: "You can delete the memories shown on your memory summary page and turn memory off by selecting Delete and turn off memory from the three-dot menu. This does not delete your past chats." And there is no single button for a total wipe. OpenAI is explicit about what that actually takes: "To fully delete something ChatGPT may know about you, you’ll need to delete every source where it appears, including past chats, archived chats, files, the memory summary, and disconnect any connected apps that may contain that information."

Correcting one wrong detail is faster than any of the above, and does not require touching the toggle at all: "You can highlight any text in the memory summary to make a specific correction." Separately, "Type what you want changed into the text box at the bottom of the memory summary and it will update accordingly."

A clean-slate prompt for one conversation, without touching any setting:

Starting now, treat this chat as if memory and past conversations do not
exist. Use only what I type below. If you are ever tempted to reference
something from an earlier chat, name it first and wait for me to confirm
it before using it.

That is a habit, not a guarantee; it works because you are instructing the model directly inside the conversation, but the account-level memory setting underneath is untouched either way. If a correction needs to stick past one conversation, starting an actual new chat after a big edit is worth doing, for the reasons covered in when a conversation goes bad, restart it.

Does turning off memory also stop ChatGPT from training on your chats?

No. They are two separate settings, documented on two separate OpenAI pages that do not reference each other. Memory, covered above, lives in Settings > Personalization > Memory and controls personalization. Training is a distinct toggle called "Improve the model for everyone", and OpenAI's Data Controls page describes what turning it off does: "Your conversations will still appear in your chat history but won't be used to train ChatGPT" (help.openai.com/en/articles/7730893-data-controls-faq, updated 18 August 2026, accessed 4 September 2026).

Neither help page mentions the other's setting. That is not a contradiction, but it is worth stating plainly because the two get conflated constantly: turning off memory does not opt your chats out of training, and opting out of training does not turn off memory. If you care about either, check both, separately.

Memory vs. custom instructions vs. projects: what's actually different?

Memory is the automatic, always-on layer: it updates itself from what you say, with no action from you, and applies across every chat by default. Custom instructions and projects are the deliberate layer: you write them once, and they stay exactly as written until you edit them. That is the practical memory vs. context distinction underneath a lot of confusion about why an answer is shaped the way it is. Memory guesses what still matters from a synthesis of everything you've said; a written instruction states what matters and does not need to guess.

OpenAI's own comparison, from the same Memory FAQ, is short and direct: legacy-style guidance continues "to allow you to provide ChatGPT with direct guidance on what you’d like it to know about you and how you’d like it to respond. For explicit information or instructions, you can add it to your Custom Instructions. For information shared via conversations, ChatGPT can remember relevant details for you." One is authored by you and static; the other is inferred by the model and moving.

This page has stayed narrowly on memory on purpose. For the full breakdown of every place ChatGPT can hold context for you, custom instructions, project instructions, files, and Custom GPTs included, with the character limits OpenAI publishes for each, see how to give ChatGPT context without repeating yourself. And memory has nothing to do with ChatGPT deciding to search the web for something current; that is a separate, explicit behavior covered in prompting ChatGPT Search.

None of this makes memory a bad feature. It makes it exactly what OpenAI's own documentation says it is: a synthesis, not a source of truth, doing its best to guess which parts of the last few months still apply today. If you would rather tell a prompt exactly what context to use than hope a synthesis got it right, that is a narrower job than turning memory off, and it is what a prompt built with the context you actually chose solves instead of guessing at. Prompt Architects's prompt enhancer builds that context into the prompt itself, on request, instead of leaving the job to whatever memory happened to keep.