Can You Ask ChatGPT to Forget Something? What Works

You typed "forget what I said about that project" and ChatGPT replied that it had forgotten. Two days later it brought the project up again, unprompted, in a completely different chat.

The reply was not a lie exactly, but it was not a confirmation either. Asking a model to forget and making a system delete stored data are two different operations, and only one of them leaves a trace you can verify.

This post separates them. What happens when you ask in chat, what actually removes information, how to tell which surface is holding the thing you want gone, and what to do about the category of context that no deletion button reaches.

The Short Answer

Asking ChatGPT to forget something inside a conversation is a request to the model, not a command to the storage layer. Sometimes it triggers a real memory update. Often it produces a polite acknowledgement and nothing else.

The reliable path is the settings panel, where saved memories are listed as discrete entries you can read and delete individually. If you can see an entry there and remove it, that removal is real. If you only have a chat message saying "understood, I have forgotten that," you have no evidence anything changed.

Treat the chat request as a suggestion and the settings list as the source of truth. Everything below is elaboration on that split.

Why the Chat Request Feels Like It Worked

Language models are built to be agreeable and fluent. When you ask one to forget something, the most natural-sounding continuation is an acknowledgement — "Got it, I won't bring that up again." Producing that sentence requires no access to storage and no verification. It is the response the training data suggests, so it is the response you get.

This is why the acknowledgement is worthless as evidence. It reads identically whether a deletion happened or not. You cannot distinguish the two cases from inside the conversation, which is exactly the situation that generates the search "chatgpt won't forget" a week later.

Worse, the acknowledgement usually is honoured — for the rest of that conversation. Within a single chat, the model has your instruction in its context window and will genuinely avoid the topic. That behaviour is real and it is temporary. It ends when the conversation ends, and the next chat has no idea the instruction ever existed.

So the pattern people experience is not random. Ask, get compliance, feel resolved, open a new chat next week, watch the thing resurface. Nothing malfunctioned. The instruction lived in a context window and context windows do not persist.

Four Places Information Actually Lives

"Make ChatGPT forget" is ambiguous because there are several distinct places something can be stored, and they have different deletion mechanics. Before you can remove something you have to work out which one is holding it.

1. The saved memory list. Discrete facts the memory feature has extracted about you. This is the only one with a per-item interface — you can read the entries and delete individual ones. When people say "ChatGPT memory," this is usually what they mean.

2. The current conversation's context window. Everything in the visible chat. The model can see all of it as long as the chat is open and it fits. Nothing is stored here in a lasting sense; the window is rebuilt each turn from the transcript.

3. Conversation history. The transcripts themselves, sitting in your account's sidebar. Deleting a conversation removes the transcript. Whether anything was extracted into the memory list before you deleted it is a separate question with a separate answer.

4. Custom instructions. The standing profile text you wrote once and forgot about. This loads into every chat, which makes it a common culprit for "it keeps knowing things I deleted." If a detail is in there, no amount of memory-deleting touches it.

The diagnostic question when something you deleted keeps resurfacing: which of these four have you actually checked? Most persistent cases are a detail sitting in custom instructions or a second memory entry phrased differently from the one you removed. The full walkthrough of each control is in ChatGPT Memory Settings: Complete Control Guide.

What Asking In Chat Can and Cannot Do

Being precise here saves a lot of wasted effort.

It can, sometimes, trigger a memory update. The memory feature can act on a request phrased as a memory operation. When it does, you generally get a visible indicator that memory was updated. That indicator is meaningful; the prose acknowledgement is not.

It can control the rest of the current conversation. Reliably. If you want the model to stop referencing something for the next hour, asking works fine.

It cannot delete conversation transcripts. No phrasing removes chats from your history. That is a separate action in the interface.

It cannot edit your custom instructions. Those are a settings field. The model reads them; it does not write them.

It cannot un-train anything. Whatever influence your data has or has not had on model training is not addressable by a chat message, and no amount of insisting changes that.

It cannot give you confirmation you can trust. This is the important one. Even in the case where deletion genuinely happened, the confirming sentence is not what tells you so. Only the settings list does.

The Verification Habit

If you want something gone, do not accept a sentence as proof. Do this instead:

1. Delete it where it lives. Find the entry in the memory list and remove it. Delete the conversation if the transcript is the concern. Edit custom instructions if the detail is standing profile text.
2. Look again. Reopen the memory list and confirm the entry is no longer there. This takes five seconds and is the only step that produces evidence.
3. Test in a fresh chat. Not a new message in the old conversation — the old conversation still has the whole transcript in its window and will look like it remembers regardless. Start a genuinely new chat and probe the topic.
4. If it resurfaces, check the other three surfaces. Almost always it is custom instructions, or a near-duplicate memory entry you did not spot, or you tested in the wrong chat.

That loop resolves the large majority of "it won't forget" cases. The remainder usually turn out to be the model inferring something obvious from what you just typed rather than recalling it — which is not a memory problem and no deletion will fix it.

When You Want to Clear Everything

Sometimes the goal is not one fact but a reset. The blunt options — clearing all memories at once, turning memory off entirely, deleting conversation history in bulk — are documented step by step in How to Make ChatGPT Forget Something, including what each one does and does not reach.

The thing worth adding here is the cost. A full clear is easy to perform and expensive to recover from, because everything useful goes with everything unwanted. The context that made ChatGPT actually helpful — how you work, your constraints, what you already tried — was accumulated slowly and disappears in one click, and there is no undo. People clear memory in frustration, get the clean slate they wanted, and spend the next month re-explaining themselves.

If you find yourself reaching for the nuclear option regularly, the problem is not that memory is too full. It is that you have no control over what goes into it, so the only tool you have is deleting all of it.

The Case Deletion Does Not Cover

There is a fifth place information lives, and it is the one that generates the most confusion: nowhere in particular.

If you spent an hour last month working through a decision with ChatGPT, that hour produced something — a conclusion, a set of constraints, a discarded approach and the reason it was discarded. Some fragment of it may have landed in the memory list, usually a thin one. The rest exists only as a transcript you will never scroll back through.

Deletion controls cannot help you here, because the problem is not that something unwanted is stored. It is that the wrong things were stored and the valuable things were not, and you had no say in either. The memory feature decides what to extract, using criteria you cannot inspect or adjust.

That asymmetry is the actual complaint underneath most forget-related searching. People do not really want their assistant to have less memory. They want it to have the right memory, and the only lever the interface offers is subtraction.

Deciding What Is Remembered Instead of Only What Is Deleted

The alternative is to keep the durable version of your context outside the assistant, in documents you write and edit, and supply it per conversation. Then "forgetting" is just editing a document, and remembering is not left to an extraction heuristic.

This is the pattern MindLock implements. In practice:

Save the conversations that produced something. In ChatGPT, Claude, Gemini, or Perplexity, press Ctrl+S (Cmd+S on Mac) at the end of a session where you decided something or solved something. The browser writes an HTML file. No extension and no API key required. There is also a Chrome extension with a one-click "Save to Memory" button, though it is developer-load only — you install it via chrome://extensions with Load unpacked rather than from a store.

Import them somewhere you control. MindLock parses saved HTML from those four platforms and stores conversations locally in IndexedDB. On the free tier nothing leaves your device. The walkthrough is in Importing Conversations.

Distil them into memory documents. Distillation turns transcripts into a profile memory plus topic memories — compact documents holding decisions, constraints, and current state rather than full dialogue. It runs locally on your GPU via WebLLM, which needs a WebGPU-compatible GPU and defaults to Llama 3.2 3B, or in the cloud via Gemini on the Pro tier, which includes 100 distillations per month. Tier details are on the pricing page.

Generate a context block and open new chats with it. Pull the relevant memories into a formatted block and paste it as your first message. See Generating Context.

Note where forgetting happens in that loop: you open the memory document, delete the paragraph, and it is gone from every future context block. No acknowledgement to distrust, no settings panel to re-check, no wondering whether a near-duplicate entry survived. The document is the state, and you can read the whole thing.

Two other properties follow from that. The documents are portable, because they are text — the same block works in Claude or Gemini. And nothing gets stored that you did not choose to store, which removes the underlying cause of most deletion requests. Memory documents are described in more detail in Memory Documents.

What This Does Not Solve

Honesty about the boundaries, since the whole point of this post is not trusting confident claims:

It does not delete anything from ChatGPT. Keeping your own memory layer is additive. If there are entries in ChatGPT's memory list you want gone, you still have to go and remove them there.

It is not automatic. Conversations are saved by you, pressing a key. Memory documents update when you re-run distillation, not continuously. Nothing watches your chats in the background.

It requires a habit. Roughly: save the sessions that mattered, re-distil when your understanding of a project has moved, paste context at the start of chats that need it. Fifteen minutes a week, not zero.

Distillation has hardware requirements on the free tier. Local distillation needs a WebGPU-compatible GPU. Without one you are on the Pro tier's cloud distillation or you cannot distil.

Bottom Line

Asking ChatGPT to forget something is a request, not a delete. It reliably controls the current conversation and unreliably touches stored memory, and the reassuring sentence you get back is the same either way.

Delete where the thing actually lives — the memory list, the conversation history, or your custom instructions — then verify by looking, and test in a fresh chat rather than the one you were just in.

If you keep needing the delete button, the real gap is that you never had control over what got saved. Keeping durable context in documents you own turns forgetting into editing, and it works the same across every assistant you use. A good place to start is Turn Your AI Chat History Into a Searchable Knowledge Base.