How to Start a ChatGPT Chat Without Memory Turned On

Sometimes you want a conversation that leaves no trace. You are asking about something personal, or testing how the model behaves with no prior context, or working on a client's problem that has nothing to do with your own projects and should not end up shaping every future answer. The memory feature is helpful right up until it is not, and what you want in that moment is one clean chat that neither reads from your stored context nor writes anything back to it.

ChatGPT supports this. What it does not do is make the boundaries obvious, and the controls involved have moved and been renamed more than once. The result is a lot of people who believe they are in a private chat and are not, and a smaller number who have switched something off globally when they only wanted it off once.

This post sorts out what "without memory" can actually mean, how to get each version of it, and — more usefully — how to verify you got it rather than assuming.

Four Different Things People Mean

Most of the confusion here is a vocabulary problem. These are four separate mechanisms with separate controls and separate consequences.

A one-off chat that is sealed in both directions. It does not draw on your saved memories, it does not add to them, and it does not show up in your conversation history. This is the temporary chat mode, and it is what most people are actually after.

Memory switched off permanently. The feature stops operating across your whole account. Nothing new is saved in any chat, and what is already stored stops being applied. Your existing entries are not deleted by this; they sit there, inert, until you either turn the feature back on or remove them.

Referencing past conversations switched off. A distinct control from the saved-memory list. One governs a curated set of extracted facts; the other governs whether the assistant can draw on your earlier chats more broadly. Turning off one does not turn off the other, and this is the single most common reason someone reports that "memory is still on after I turned it off".

Not being signed in at all. A logged-out session has no account to attach memory to. This is the strongest separation available and also the least convenient, and it is a different thing from a private browsing window — an incognito tab that you then sign into is an ordinary signed-in session with an ordinary memory store attached.

Pick which one you want before touching any settings, because the cheapest of them is temporary chat and it is reversible in a way the others are not.

Temporary Chat: What It Actually Does

The mechanism is a per-conversation mode rather than a setting you leave on. You start a chat in temporary mode, use it, and close it, and your defaults are untouched for the next chat.

Inside one, three things hold. Your saved memories are not supplied to the model, so it starts from zero about you. Nothing from the conversation is extracted into memory, so your stored context is not polluted by whatever you did in there. And the conversation does not appear in your history sidebar, so it cannot be scrolled back to later.

That combination is what makes it the right default for one-off work. You are not making a decision that leaks into next week, and you are not spending anything to clean up afterwards.

Where the control lives: in the chat interface rather than in the settings screen, near the model selector or in the menu attached to a new conversation. I am deliberately not giving you a click path, because the exact position has changed repeatedly across product versions and a stale one sends people hunting for a control that has moved two menus over. Look for the wording on your own screen. A temporary conversation is normally also marked as such once you are in it, which is the thing to actually rely on.

What ends it: closing or navigating away from it. It is not a mode that persists — this is a feature, but it also means the next chat you open is a normal one unless you deliberately start another temporary one. Assuming stickiness is how people end up saving something into memory that they meant to keep out.

What a Temporary Chat Does Not Do

Worth being precise about, because the mode is sometimes described as though it were anonymity, and it is not.

It is not logged out. The request still goes to OpenAI over your authenticated session. Your account is still the account making it.

It is not zero-retention. Temporary conversations are generally retained for some period for abuse monitoring and safety review even though they are excluded from your history and from memory. The specific retention period is a policy detail that changes, so read the current wording on OpenAI's own help pages rather than trusting any number quoted in an article, including this one.

It does not retroactively remove anything. Facts already in your memory list stay there. If your goal is to get rid of something that was stored earlier, temporary chats do nothing for you at all and you want How to Make ChatGPT Forget Something instead.

It does not make the model a different model. Custom instructions and account-level personalisation may still be in effect depending on how those are scoped. If you are testing the model's unprimed behaviour rather than protecting privacy, that distinction matters, and the honest test is a logged-out session.

Turning Memory Off Properly

If you want the account-wide version rather than the per-chat one, the thing to know is that there is more than one switch and they are independent.

The saved-memory control governs the list of extracted facts — the entries you can open and read and delete individually. Switching it off stops new entries being written and stops existing ones being supplied.

The chat-history-referencing control is separate and governs whether earlier conversations can be drawn on at all. It is entirely possible to have the first off and the second on, at which point the assistant still appears to know things about you and you conclude the setting is broken. It is not broken; you switched off one of two mechanisms.

Check: open your settings and read both controls on your own screen, not against a screenshot from an article. Then confirm the saved list is in the state you expect — off is not the same as empty, and an account with the feature disabled can still hold dozens of stored entries waiting to become live again the moment it is re-enabled. The full tour of the controls and what each one governs is in ChatGPT Memory Settings: Complete Control.

If you want it genuinely gone rather than paused, disabling is not sufficient and you need to clear the entries themselves. Deletion has no undo, so copy the list into a text file first. That procedure is set out in Make ChatGPT Forget Everything About You.

Verifying It Rather Than Assuming It

This is the part that gets skipped, and it is the only part that actually settles anything. Two minutes.

Test the read path. In the chat you believe is memoryless, ask about something distinctive that is definitely in your stored memory — a project name, a preference, a detail about your work that you know was saved. A correct answer means the chat is drawing on memory and is not sealed. A genuine blank means the read path is closed.

Be careful constructing this one. Ask about something the model could not plausibly guess or infer from your phrasing. "What do you know about me?" is a weak test, because a well-behaved assistant with no memory will still produce a hedged, agreeable non-answer that is easy to misread as a blank.

Test the write path. State something invented and distinctive in the chat — an unusual codename, a fake deadline, a preference you do not hold. Close the chat. Then open an ordinary new conversation and ask about it, and also look at your saved-memory list directly. If neither knows, nothing was written.

Test that you are in the account you think you are in. Memory is per-account, and a second identity, a work profile or a different browser has its own store. An empty store looks exactly like a disabled feature. Read the email address in the account menu instead of assuming.

Run those three once, after any settings change, and you will never again be in the position of half-believing a chat is private.

Choosing Between Them

Short version of the decision.

Temporary chat for anything one-off: sensitive questions, a client's problem, testing behaviour, a topic you do not want shaping future answers. It is reversible, scoped, and costs nothing. This should be your reflex.

Memory off account-wide if you have concluded the feature is a net negative for how you work — if it keeps surfacing stale context, or you dislike the unpredictability, or you would simply rather supply context deliberately. This is a considered decision, not a fix for a single chat.

Referencing past chats off, memory on if you want a curated set of stable facts about you and nothing else. This is underrated and is the setting most people actually want when they say memory is too noisy.

Logged out when the concern is genuinely about association with your account rather than about conversational cleanliness. The privacy reasoning behind that choice, and where its limits are, is in Private AI Memory and Data Sovereignty.

The Thing Nobody Mentions

Turning memory off works. It also removes the only continuity you had, and the reason people turn it back on within a week is that the alternative is re-explaining yourself at the start of every conversation.

That is the real trade being made, and it is a bad trade presented as a binary. The built-in feature gives you two options — on, where an opaque process decides what to store about you and applies it everywhere including where you did not want it, or off, where you start from zero every time. Neither is "context I control", which is the thing you actually wanted.

The gap is structural rather than a missing setting. Memory as implemented is automatic, invisible in its selection criteria, scoped to one vendor, and all-or-nothing per account. What most people want is the opposite on every axis: explicit, readable, portable, and chosen per conversation.

Memory You Supply on Purpose

That inversion is what MindLock does. Nothing is captured in the background and nothing is applied without you asking for it.

You save the conversations that were worth keeping with Ctrl or Cmd plus S in ChatGPT, Claude, Gemini or Perplexity, or with the one-click button from the sideloaded Chrome extension, then upload the saved page. Distillation turns them into memory documents — a profile memory plus topic memories — that you can open and read in full. Starting a new chat, you pick which memories are relevant, generate a context block, and paste it in. A chat you do not paste into is a chat with no memory, with no setting to remember to flip.

That is the part that solves this post's problem: you get the clean-slate conversation and the continuous one from the same setup, chosen per chat, and neither requires trusting a toggle you cannot inspect. How conversations get in is in Importing Conversations.

Two honest caveats. It is manual by design — saving is a keystroke, and a memory document only reflects your last distillation rather than updating continuously. And local distillation needs a WebGPU-capable GPU; without one you would use the cloud option on the paid tier, and the split is on the pricing page.

The Short Version

For a single conversation that neither reads nor writes memory, use temporary chat. It is per-chat, it does not persist to the next one, and it leaves your account settings alone. It is not the same as being logged out and it is not zero-retention.

For the account-wide version, remember there are two independent controls — the saved-memory list and the referencing of past conversations — and switching off one leaves the other running. That is almost always the explanation when memory seems to persist after being disabled.

Then verify instead of assuming. Ask about something you know is stored to test the read path, plant something invented and check afterwards to test the write path, and confirm which account you are signed into. Two minutes, and it is the difference between a private chat and one you merely hoped was private.