Claude Memory: How to Control What Claude Remembers

Claude memory is switched on by default on the Free, Pro and Max plans, and most people who have it running have never opened the panel that controls it. Since 25 August 2026 it also carries what it learns between ordinary chats and Claude Cowork, which means a detail you mentioned once in a throwaway conversation can resurface inside a work task days later. The setting that would keep those two apart does not exist.

This guide is the control panel rather than the explainer. It covers where the settings actually live, what each switch does, and which action is permanent. It also covers how to remove a single stored fact instead of wiping everything. And it covers the thing almost everyone gets wrong: deleting a conversation does not delete what Claude learned from it.

The Key Takeaways

• On by default: memory is on for Free, Pro and Max, and off by default for Team and Enterprise, where an owner has to enable it.
• Two paths: the settings sit at Settings > Memory in the current interface, or Settings > Capabilities in the older one.
• Pause is reversible, Reset is not. Reset permanently deletes every memory including project memories, and Anthropic states it cannot be undone.
• Deleting a chat is not deleting a memory. Entries generated from a conversation stay put after the conversation is gone.
• Legacy memory export closes 9 September 2026. After that the old export route is gone.

What Claude Memory Actually Stores

Claude memory is not a transcript of your conversations. It is a set of individual entries that Claude writes as you talk, each one a short fact or preference rather than a saved chat. Anthropic groups them under Topics, and the list is readable, editable and deletable item by item.

In practice the entries cluster around who you are and how you work: your role, the projects you keep returning to, the tone you prefer, the tools and languages you use. That is what makes the feature useful, and it is also why it is worth auditing.

A preference you expressed once during a bad week can quietly steer answers for months.

Projects each keep their own separate memory space, so context from one project does not leak into another. That separation is real and it is the one boundary Claude does enforce cleanly.

Where the Claude Memory Settings Live

The single most common reason people give up on this is that they look in the wrong place. There are two valid paths depending on which interface your account is on, and Anthropic documents both.

The Two Settings Paths

Which path you have tells you which version of memory you are on. If you see Settings > Memory, you are on the current experience. If memory sits under Settings > Capabilities, you are still on the legacy one, and that distinction matters for the export deadline further down.

Memory works on the web, in Claude Desktop and in the mobile apps. The settings are account-level, so changing them in one place changes them everywhere you are signed in.

Every Switch and What It Does

How to Turn Claude Memory Off, Pause It or Reset It

These three sound interchangeable and are not. Picking the wrong one either leaves your stored history intact when you wanted it gone, or destroys it when you only wanted a break.

Stop New Memories Being Written

Open Settings > Memory and toggle off Generate memory from chats. Everything already stored stays exactly where it is and Claude can still draw on it. This is the right choice if the existing memory is useful but you are about to have a run of conversations you would rather not have absorbed.

Pause Memory

Pause goes one step further. Claude keeps the entries but stops both using them and adding to them, so the account behaves as though memory were off while nothing is lost. It is the reversible option, and it is the one to reach for when you are not sure yet.

Reset Memory

Reset permanently deletes every memory on the account, including project memories. Anthropic is explicit that once you select it and confirm, this cannot be undone. There is no archive, no grace period and no export prompt on the way out.

If there is any chance you will want the context later, pause instead.

How to Edit or Delete a Single Memory

Wiping everything to remove one wrong fact is the mistake worth avoiding, and it is unnecessary. Open Settings > Memory > Topics and you get the full list of what Claude has stored. Open any entry to correct it or remove it on its own. The change applies to future conversations rather than rewriting past ones.

Correcting an entry is usually better than deleting it. Say Claude has recorded that you work in a language you have since moved off. Editing the entry replaces that assumption outright. Deleting it leaves a gap, and Claude will refill the gap from whatever you happen to say next. The same applies to a stale job title or a project that finished months ago.

It is worth doing this on a schedule rather than when something goes wrong. Imported or long-standing context drifts as your work moves on, and a memory list nobody has read in three months is where confidently wrong answers come from. Our guide to ChatGPT memory and how to delete it covers the same audit habit on the OpenAI side, where the storage model is auto-managed and rather less transparent.

Memory Now Follows You Into Cowork

This is the change most guides have not caught up with. On 25 August 2026 Anthropic unified the memory systems behind ordinary chat and Claude Cowork, so that context learned in one surfaces in the other. Anthropic's release notes put it plainly: memory now works across chat and Cowork in the cloud.

For anyone using Cowork for real work this is mostly welcome, because it removes the tedious business of re-explaining your setup every time you move between a conversation and a task. According to TechCrunch's report on the change, it was one of the most requested fixes for Cowork.

The catch is the one that belongs in a control guide. There is no setting that keeps chat memory and Cowork memory separate while memory is switched on. It is one shared store or none, so the only way to keep a personal conversation out of a work context is to stop the entry being written in the first place.

One store, or nothing. If Cowork is part of your daily routine, our complete Claude Cowork guide covers how the two surfaces now fit together.

What Deleting a Chat Does Not Do

Here is the assumption that catches almost everyone. Deleting a conversation does not delete the memory entries Claude generated from it. Anthropic documents this directly: when conversations are deleted or expire, the related memory entries are not automatically removed.

They are two separate stores in two separate places. The delete button in your chat list only touches one of them. Someone who clears a sensitive conversation and assumes the matter is closed has cleared the transcript and left the derived facts behind.

To remove it properly, go to Settings > Memory > Topics and delete the entry as well. Both, in that order.

Sensitive Topics and What Claude Never Saves

By default Claude keeps a defined set of categories out of memory entirely: health, race, ethnicity, religion, politics and gender identity. They are stored only if you deliberately switch on Include sensitive topics in memory, and there is rarely a good reason to.

A second tier is never stored at all regardless of that setting. Anthropic lists government ID numbers, criminal history, financial account numbers and immigration status as excluded outright, with no opt-in available. That is a sensible default and a better position than several competitors take. It is still not a reason to paste any of it into a chat window. For the wider picture on what these systems retain and how to limit it, see our 2026 guide to stopping AI training on your data.

The Legacy Memory Export Deadline

If your account predates the current entry-based system, the older rolling memory summary can still be exported, but only until 9 September 2026. After that date the legacy export route closes and whatever was in the old format is no longer retrievable through it.

This only matters if you have been using Claude long enough to have accumulated something in the old system and you want a copy outside Anthropic. Legacy users find it under Settings > Capabilities, then View and edit your memory. On the current experience the same option sits under Settings > Memory. It is the kind of deadline that is trivial to meet now and impossible to meet later. Note that this is a different thing from bringing memory in from another assistant, which our guide to switching from ChatGPT to Claude covers.

The Five Minute Claude Memory Audit

If you do nothing else after reading this, do these five things once. Open Settings > Memory > Topics and read the whole list, which is usually shorter and more revealing than people expect. Correct anything out of date rather than deleting it. Delete anything you would not want surfacing inside a Cowork task now that the two share a store. Confirm Include sensitive topics in memory is off unless you turned it on deliberately. Then decide whether Generate memory from chats should stay on at all for the way you actually use the account.

That last question deserves a real answer rather than a default yes. Memory earns its place if you do sustained work on a small number of long-running projects. It earns much less if your usage is a scatter of unrelated one-off questions. The entries it writes from those are noise, and later answers still get steered by them.

The Verdict

Claude's memory controls are better than most, mainly because the stored entries are readable and individually editable instead of being a black box you can only wipe. The transparency is the feature. What it asks in return is that you actually look at the list occasionally, which almost nobody does.

Two things deserve a decision today rather than eventually: check whether anything in Topics should not be crossing into Cowork, and export legacy memory before 9 September 2026 if that applies to you. Everything else can wait for your next audit. Still weighing Claude against the alternatives? Our full guide to Claude AI covers the rest of the picture. Anthropic's own memory documentation is the reference for any label that changes after publication.

FAQ