November 8, 2026•12 min

How to Make Claude Forget a Conversation (What Works)

You told Claude something you wish you had not. Maybe it was a detail about a client, an old version of a plan you have since abandoned, or a personal fact that has no business turning up in future answers. Now you want it gone, and the obvious move, typing "forget what I just said", does not feel like it should be enough. It is not, but the reason is more useful than the frustration: "forget" means several different things in Claude, and each one lives in a different place with a different off switch. This post separates them, tells you which control applies to which, and gives you a short test to confirm the thing you wanted gone is actually gone.

Why "Forget" Is Five Questions, Not One

When people search whether they can make Claude forget a conversation, they are usually asking one of five things without realising they are different:

  1. Can I make Claude stop using something inside the chat I am in right now?
  2. Can I remove a conversation from my history?
  3. Can I stop something from showing up in other chats, particularly inside a Project?
  4. Can I stop an account-level memory or past-chat feature from using it?
  5. Can I make Anthropic not keep it at all?

These are answered by different mechanisms. The first is about the context window of a single conversation. The second is about your chat list. The third is about Project instructions and attached documents. The fourth depends on which features are enabled on your account. The fifth is about retention, which is a policy question rather than a product setting you can verify from the chat window.

The common mistake is solving the wrong one. Deleting a chat does nothing to a Project document that contains the same fact. Asking Claude to disregard something does nothing to whether the conversation is stored. Each layer needs its own action, and the rest of this post walks through them in order of how often they are the real problem.

If you want the background on what Claude carries between chats in the first place, it is laid out in Does Claude Remember Previous Conversations?. The short version: a new conversation starts with an empty context window, and anything that appears to be remembered has been deliberately put somewhere that gets re-supplied.

Layer 1: The Conversation You Are In

Inside a single conversation, Claude works from the transcript. Every message you have sent and every reply it has given is part of what the model reads before producing the next answer. That has a direct consequence for forgetting: you cannot remove text from the transcript by asking.

When you write "forget what I said about the budget", what actually happens is that a new instruction gets added to the conversation. The original budget figure is still there, a few messages up. The model can follow your instruction to ignore it, and often will, but the information has not left. It is still being read on every turn, and in a long conversation, instructions given late can carry less weight than you expect against material that has been discussed at length.

So for the chat you are in, there are three realistic options, from weakest to strongest:

Ask Claude to disregard it. Fine for low-stakes corrections, such as a figure you mistyped. Be specific: "The budget figure I gave earlier was wrong. Ignore it entirely and use 40,000 from here on." Vague instructions to forget tend to be followed vaguely.

Edit the earlier message, where the interface lets you. Editing a previous message and resubmitting it takes the conversation down a new path from that point, so the replies built on the original version are no longer the ones the next answer follows. This is a cleaner fix than an instruction to ignore something, because the corrected text is what the conversation continues from. Check how your own interface handles edited messages and earlier versions before relying on it for anything sensitive.

Start a new conversation. This is the only way to guarantee the current context window no longer contains the thing. Before you leave, write down what you want to carry forward, leaving out what you want forgotten, and paste that into the new chat. A short handover like this is described in detail in A Context Brief Template That Re-Briefs Any AI.

The third option is usually the right one when the thing you want forgotten has influenced many replies. A wrong assumption that shaped twenty messages is easier to leave behind than to argue out of the transcript.

Layer 2: Your Chat History

Your conversation list is the second place people look. Deleting a conversation from it does two things, and it is worth being precise about both.

It removes the conversation from your view. You will not see it in the sidebar, search for it, or reopen it.

It does not change what any other conversation already contains. Since a new Claude conversation does not start with previous chats loaded into it, a chat you never referenced in another chat was not influencing that other chat to begin with. If you copied part of it into a Project document or pasted it into another conversation, those copies are separate and are untouched by the deletion.

That second point surprises people in both directions. Some expect deleting a chat to undo its influence everywhere, and are confused when a Project still knows the fact. Others hesitate to delete because they assume it will break other conversations, when it will not.

Deleting from your history is also not the same as deletion from the provider's systems. What happens to deleted conversations on Anthropic's side, and on what timeline, is set out in Anthropic's privacy documentation, and that is the only source to trust on it. A blog post, including this one, is not in a position to tell you how long anything is retained on your account today. More on that in Layer 5.

Layer 3: Projects, Instructions, and Attached Documents

If something keeps coming back in new chats, a Project is the most likely reason. A Claude Project is a workspace with instructions and attached documents, and every conversation started inside it begins with that material available. It is the main supported way for context to persist across Claude conversations, which also makes it the main place unwanted context persists.

Work through it in this order:

  1. Read the Project instructions. Instructions are easy to forget about once written, and they apply to every chat in the Project. An old constraint or a personal detail written there months ago will be applied every time.
  2. Check each attached document. Open the files, not just their names. The fact you want gone is often inside a spec, a set of notes, or an export you attached early on and never reopened.
  3. Remove or edit the source. Delete the document from the Project, or replace it with a corrected version. Editing a local copy of a file does not change the copy already attached; the attached version is the one Claude reads.
  4. Start a fresh chat in the Project to confirm. Existing conversations in the Project still contain whatever they already discussed. The change applies to what new conversations start with.

The related question, whether chats inside the same Project can see each other, comes up constantly and has its own answer in Does Claude Remember Chats in the Same Project?. The practical point for forgetting is that a Project's persistent material is what you put in its instructions and documents, so that is where you remove things.

Remember too that Projects are isolated from each other. If you copied the same notes into several Projects, you need to clean each one. There is no global setting that reaches into all of them at once.

Layer 4: Account-Level Memory and Past-Chat Features

Memory and conversation-history features are among the fastest-moving parts of every AI assistant, and Claude is no exception. Capabilities that reference past conversations or build up memory over time ship, change names, and roll out to some plans and accounts before others.

That means any precise instruction here, such as "toggle this switch in this menu", is likely to be out of date by the time you read it. What holds regardless:

  • Open your Claude settings and look. Find whatever controls exist for memory, chat history, or referencing past conversations on your account. That takes less time than reading another article about it, and it is authoritative for your account in a way no article can be.
  • If a feature draws on past chats, deleting the source chat is the relevant action, along with any controls the feature itself provides for viewing, editing, or turning it off. Read the description shown next to the setting; it is the most current statement of what that feature does.
  • If no such feature appears in your settings, this layer does not apply to you. In that case, persistence comes only from Layers 2 and 3.

The broader overview of what Claude currently persists, and how Projects behave, is in What Claude AI Actually Remembers in 2026. Treat it the same way as this section: the structural parts are stable, the feature list is a snapshot.

Layer 5: Retention Is a Different Question

The previous four layers are about whether Claude will use something. Retention is about whether it is stored, and it is the question that matters most if the information was genuinely sensitive.

These are not the same. A conversation can be absent from every future chat and still be retained for some period under the provider's policy. Conversely, a fact can be stored and never used again. Mixing the two leads to false confidence: someone deletes a chat, sees that new chats do not know about it, and concludes it no longer exists anywhere.

For retention, the only reliable sources are Anthropic's privacy policy, its help documentation, and the privacy controls shown in your own account. Read them for your situation, particularly if you are on a work or team account, where an organisation's settings may also apply.

The more durable lesson is preventive. If something should not be stored by a hosted assistant, the safest approach is not to put it there. Replace names with placeholders, summarise rather than paste, and keep the genuinely sensitive part of your notes somewhere you control. That is less convenient than cleaning up afterwards, and far more reliable.

A Quick Decision Guide

Find your situation and take the matching action:

  • "I mistyped something in this chat." Correct it explicitly, or edit the message where the interface lets you.
  • "This chat went down a wrong path and the assumption is everywhere." Write a clean handover without the wrong part and start a new conversation.
  • "I want this conversation out of my list." Delete it from your chat history.
  • "New chats in a Project keep using an old fact." Edit the Project instructions and remove or replace the attached document that contains it.
  • "New chats outside any Project seem to know something from before." Check your settings for memory or past-chat features and their controls.
  • "I am worried the content is stored at all." Read Anthropic's privacy documentation and your account's privacy settings, and in future, keep that material out of hosted chats.

Most real cases are a combination. A detail you regret sharing might be in the current chat, in a Project document you attached, and in two older conversations. Clean each layer it touches.

How to Check That It Worked

Do not assume forgetting worked because you took an action. Test it, and test it from a fresh conversation, because the old conversation still contains its own transcript.

  1. Open a brand-new chat outside any Project. Ask a question that would only be answered correctly with the forgotten information. Phrase it neutrally, so you are not supplying the answer in the question.
  2. Open a brand-new chat inside each relevant Project. Ask the same question. This catches facts living in instructions or attached files.
  3. Ask directly what Claude knows about the topic. "What do you know about my budget for this project?" is a reasonable probe. An answer that recites the old figure tells you where to keep looking.

If the information turns up, the answer tells you which layer still has it. Appearing only inside a Project points at its instructions or documents. Appearing outside Projects points at an account-level feature, if your account has one. A fuller method for probing what any assistant actually retains is in How to Test What Your AI Assistant Remembers.

How This Compares to ChatGPT

ChatGPT's memory works differently, with a list of saved facts that can be viewed and edited, alongside its own chat history settings. The layered thinking is the same, though: what is in the current chat, what is in history, what is in saved memory, and what is retained are separate questions with separate controls. If you use both assistants and want the ChatGPT side covered too, the equivalent walkthrough is How to Make ChatGPT Forget Something.

The thing neither assistant does is reach across to the other. Forgetting something in Claude has no effect on a ChatGPT conversation where you mentioned the same thing, and the reverse is also true. If a fact has spread across tools, you clean it up in each one.

Where MindLock Fits

MindLock is a separate memory layer that sits outside Claude, so it has its own forgetting controls, and they are simpler because the data is yours to manage directly.

You save the Claude conversations you want to keep by pressing Ctrl or Cmd plus S on the conversation page and uploading the saved file on the Conversations page, or with the one-click Save to Memory button in the Chrome extension, which is loaded unpacked in developer mode rather than installed from a store. Nothing is captured automatically, so a conversation you never saved is not in MindLock at all. The import steps are in Importing Conversations.

Forgetting in MindLock works at two levels:

  • Conversations. A saved conversation can be deleted from its page in the app.
  • Memory documents. Distillation turns saved conversations into a profile memory and topic memories. These are documents you can open, edit, and delete. If a memory contains a detail you want gone, edit it out or delete the document. Memory documents do not update on their own; they change when you edit them or re-run distillation. More detail is in Memory Documents.

One practical consequence: if you delete a conversation but keep a memory document distilled from it, the distilled content is still in that document until you edit it or distil again from your remaining conversations. Clean both if the detail matters.

On the free tier, conversations and memory documents are stored in your browser's local storage on your own device, and local distillation runs on your GPU through WebGPU, which needs a compatible GPU. The Pro tier adds cloud distillation and Firebase cloud sync; the details are on the pricing page. Keeping your long-term context in a place you control is also the cleanest answer to Layer 5: material that never went into a hosted chat does not need to be forgotten by one.

The Short Version

Telling Claude to forget something adds an instruction; it does not remove anything. To actually make Claude stop using information, act on the layer where it lives: start a new chat or edit the message for the current conversation, delete the chat from your history, clean Project instructions and attached documents, and check your account settings for any memory or past-chat features. Retention is a separate question answered only by Anthropic's own documentation. Then test from a fresh chat, inside and outside your Projects, to confirm it worked. The most reliable forgetting of all is not sharing the sensitive part to begin with.