Tell ChatGPT Not to Remember Something: 4 Reliable Ways

You are about to share something with ChatGPT that you do not want following you around: a medical detail, a one-off client name, a draft you are not proud of, or a question you would rather not see reflected back in a month. The question is whether you can tell it, in advance, to leave this one out. The search data suggests plenty of people are asking exactly that, sometimes in a hurry and with the typos to prove it.

The honest answer is that you can, but asking in plain words is the weakest of the available options and the only one you cannot fully verify ahead of time. There are stronger tools, they are already in the product, and they work before anything is stored rather than after. This post walks through each option from most to least reliable, explains what each one actually prevents, and shows you how to confirm afterwards that nothing was kept.

Why "Before" Is Different From "After"

Most advice on this subject is about forgetting: deleting a saved memory, deleting a chat, clearing everything. That is useful, and it is covered in Can You Ask ChatGPT to Forget Something?. But deletion is cleanup. It assumes the thing was stored and then removes it.

Preventing storage is a different problem with different tools. The appeal is obvious: if a fact is never written to your saved memory list, there is nothing to find, nothing to delete and nothing to resurface in an unrelated conversation later. The catch is that you have to choose the right tool before you type, because most of these options cannot be applied retroactively to a conversation that is already under way.

It also helps to be clear about what "remember" means here. There are, broadly, three ways something you say can reach a future conversation:

1. The saved memory list. Facts extracted about you and kept in a list you can view and edit in settings.
2. Chat history referencing. Depending on your plan and settings, ChatGPT may be able to draw on past conversations themselves rather than only on the saved list.
3. The conversation record itself. The chat sits in your history, where you, or anyone with access to your account, can open it.

A good "do not remember" strategy deals with all three. Telling the model in words mostly addresses the first one, and only partially.

Option 1: Use a Temporary Chat

For a single sensitive conversation, a Temporary Chat is the strongest tool you have. It is designed specifically for the case where you want a conversation that does not feed your memory and does not sit in your history.

In practice that means:

• It does not draw on your saved memories.
• It does not add new entries to your saved memory list.
• It does not appear in your chat history list.

Read the provider's own description of the feature before relying on it for anything important, because the precise details, including any retention for safety review, are set by the provider and can change. The point here is structural: a Temporary Chat stops the memory problem at the door, which no instruction typed into a normal chat can do as reliably.

The trade-off is that it starts with no memory at all. Whatever ChatGPT normally knows about your preferences is not applied, so you may need to give more context in the first message. For a sensitive one-off, that is usually exactly what you want. A fuller walkthrough of how it behaves is in ChatGPT Chat Without Memory.

One practical warning: start the Temporary Chat before you type the sensitive part. Opening a normal chat, sharing the detail and then switching does not move what you already wrote out of the normal conversation.

Option 2: Turn Memory Off for a While

If you are about to do a run of sensitive work across several conversations, toggling a temporary chat each time gets tedious. The alternative is switching off memory in settings for the duration.

ChatGPT's memory controls have been renamed and reorganised more than once, so the labels on your screen may not match any article, including this one. Look for two kinds of control:

• One that governs whether ChatGPT uses and updates your saved memories.
• One that governs whether it can reference your chat history.

Switching both off means new conversations neither read from nor add to either source. Your existing saved list is not deleted by turning the feature off; it simply is not used while the setting is off. That is useful, because it means you can pause and resume without losing what you had built up.

Two things to remember. First, the conversations themselves still appear in your history unless you delete them, so this option prevents memory, not record-keeping. Second, it is easy to forget you switched it off. Put a note somewhere you will see it, or set a reminder, so you do not spend the next fortnight wondering why ChatGPT has stopped knowing your preferences. The full map of what each control does is in ChatGPT Memory Settings: Complete Control.

Option 3: Ask in Plain Words

This is what most people try first, and it is the least reliable. You can write something like "Do not save anything from this conversation to memory" or "Do not remember the following".

It often works, in the sense that the model will frequently comply and not write a memory entry. But there are three reasons not to rely on it for anything that genuinely matters:

Compliance is not a guarantee. The model decides what to extract based on the conversation as a whole. An instruction early in a long chat may carry less weight by the time a memory-worthy detail comes up much later, especially if the conversation has grown large enough for its earliest parts to fade.

The instruction covers saved memories, not history. Even when it is honoured, the conversation still sits in your history. If your account can reference past chats, the instruction does not obviously stop that.

You cannot see the decision being made. There is no indicator that tells you "nothing was saved because you asked". You only find out by checking afterwards.

If you do use this approach, make the instruction specific and put it immediately before the sensitive detail rather than at the top of the chat: "The next message contains details I do not want saved to memory. Do not store any of it." Then check, as described below.

Option 4: Keep the Detail Out Entirely

Sometimes the most reliable option is not to share the identifying part at all. You can often get the help you need with the specifics removed:

• Replace real names with roles: "a client", "my manager", "the supplier".
• Replace exact figures with ranges or placeholders if the precise number does not change the advice.
• Describe a situation in general terms rather than pasting the original document.

This is not always possible. Some questions only make sense with the real details. But it is worth asking yourself whether the specific identifying fact is actually needed for the answer. If it is not, leaving it out is the only option that works regardless of settings, plans and product changes.

How to Check That Nothing Was Kept

Whichever option you used, verify it. It takes a couple of minutes.

1. Open your saved memory list in settings and reload it. Look at the newest entries. If anything from the sensitive conversation appears, delete it.
2. Start a brand-new normal chat. Not a branch of the sensitive conversation, and not a Temporary Chat.
3. Ask a neutral question that would naturally draw on the detail. Do not name the detail itself, because then you are providing it rather than testing for it. If you shared a health condition, ask for general meal-planning ideas and see whether the answer assumes the condition.
4. Check your history. If you used a normal chat and want no record, delete the conversation as well.

If the fact shows up in step 3 despite not appearing in the saved list, it reached the new chat some other way, usually through chat history referencing or through your custom instructions. How to trace which source is responsible is covered in Make ChatGPT Forget Everything About You.

Common Situations and the Right Tool

To make this concrete, here is how the options map onto typical cases.

A one-off sensitive question. Temporary Chat. Start it before you type anything.

A week of work on a confidential project. Turn memory off for the duration, delete the conversations afterwards if you do not want them in history, and turn memory back on when you are done.

A normal conversation where one detail slipped in. Too late for prevention. Check the saved list and delete anything that was extracted, then delete or keep the conversation depending on whether you want the record.

A preference you do not want generalised. For example, you asked for a very formal tone once for a specific letter and do not want every future reply to be formal. Say so in the conversation ("this is a one-off, not a general preference"), then check the saved list afterwards. This is a case where the plain-words approach is usually adequate, because the stakes are low and the check is quick.

Testing what ChatGPT already knows. Before deciding what to prevent, it can help to find out what is currently stored and applied. A structured way of doing that is in Test What Your AI Assistant Remembers.

Why This Is Harder Than It Should Be

It is reasonable to find all of this more effortful than it ought to be. The underlying issue is that the memory features are designed to be helpful by default, quietly picking up useful facts, and quiet extraction is exactly what you do not want when the material is sensitive.

The fix is not to distrust memory entirely. It is useful for preferences and long-running context. The fix is to decide, before a conversation starts, which category it falls into, and to use the tool that matches. Sensitive and one-off: temporary. Sensitive and ongoing: memory off for the duration. Ordinary: leave it on and prune the list occasionally.

Keeping Context You Choose, in a Place You Control

The other side of this is the context you do want to keep. If you have been using ChatGPT's memory as the only place your preferences and project details live, every decision about what to prevent or delete is also a decision about what you might lose.

MindLock takes a different approach: you decide which conversations to keep, and the result is text you can read and edit. You save a conversation with Ctrl or Cmd plus S in ChatGPT, Claude, Gemini or Perplexity, or with the "Save to Memory" button from the Chrome extension, which is loaded unpacked in developer mode rather than installed from a store. Uploaded conversations are distilled into memory documents, a profile memory plus topic memories, which you can inspect, edit and delete directly. More detail on what those documents contain is in Memory Documents.

On the free tier, storage is local to your browser in IndexedDB, and distillation runs on your own GPU through WebGPU when your hardware supports it. Cloud sync and cloud distillation are part of the Pro tier, described on the pricing page. Nothing is captured automatically: a conversation is only included if you save and upload it, which is the point. A conversation you never save is simply never there.

The Short Version

You can tell ChatGPT not to remember something, but asking in plain words is the weakest option and cannot be verified until afterwards. For a one-off sensitive conversation, start a Temporary Chat before you type, which keeps the conversation out of memory and out of your history list. For a run of sensitive work, switch off the memory and chat-history controls in settings for the duration, then switch them back on.

If the identifying detail is not needed for the answer, leave it out entirely; that is the only method that works regardless of settings. Whatever you choose, check afterwards: reload the saved memory list, test in a fresh normal chat with a neutral question, and delete the conversation if you want no record of it.