Claude memory: what it keeps, how to control it, and when to switch it off

A finance analyst mentions, in passing, that her company's fiscal year starts in February. A month later she opens a new chat and asks for "Q1 revenue growth". Claude uses February to April without being told. That is memory working.

Three months after that, the company moves to a calendar year. Nobody tells Claude. Every "Q1" it computes is now wrong, and nothing in the answer says why. That is memory working too.

Claude's memory saves information "as a set of individual topics as you chat", rather than writing a summary after each conversation ends (Anthropic Help Centre, checked October 2026). You can also ask directly: "remember that our reporting currency is GBP". It sits alongside two related features that are easy to confuse:

Memory reached Team and Enterprise in September 2025, then Pro and Max in October 2025, alongside incognito chats and project-scoped memory (Anthropic, October 2025). The help centre now lists it as on by default for Free too.

The controls are in Settings > Memory, and they differ more than their names suggest:

• Generate memory from chats. The main switch. Off means Claude neither creates nor uses memories.
• Pause memory. Stops Claude using or creating memories but keeps what's saved. Use it for a week of work for a different client.
• Reset memory. "Permanently deletes all memories including project memories." There is no undo.
• Search and reference chats. Controls chat search separately from memory.

To manage individual memories, open Settings > Memory, where they're listed under Topics. Select one to read it, edit it, or delete it. From inside a chat you can tell Claude what to remember, change or forget; the change applies from your next conversation.

For one conversation that shouldn't touch memory, you have two choices. Turn off "Memory" in the "+" menu before you send the first message, or start an incognito chat (the ghost icon), which isn't saved to history, memory or chat search. On Team and Enterprise, your organisation's retention settings still apply to incognito chats.

Projects get their own memory. Inside a project, Claude uses that project's memory plus your account memory, never another project's. To wall a project off completely, turn off "Use account memory" on the project's page. For client work, we'd do that by default.

Memory is one more thing an answer can quietly rest on. Count connects to Claude and labels each claim by what supports it, so a figure with no source in this conversation stands out instead of reading like the rest.

Anthropic has drawn some lines (Anthropic Help Centre and blog, checked October 2026):

• Sensitive topics such as health, politics and religion aren't stored by default. You can opt in under Settings > Memory.
• Government ID and financial account numbers are never saved.
• Deleting a chat does not delete memories made from it. The help page is explicit: "related memory entries generated from it won't be removed." If a chat contained something you'd rather Claude forgot, delete the memory topic as well as the chat.
• Memory is included in data exports, so it's portable and auditable.

Before rolling memory out to Pro and Max, Anthropic says it tested whether memory could reinforce harmful patterns, make Claude over-accommodating, or help bypass safeguards, and adjusted the feature in response. That's reassuring about safety. It says nothing about whether a remembered fact is still true.

Our rule of thumb: let memory hold how you like to work, and keep what is true about the business somewhere you can see it.

Where it helps

• Style and format. British spelling, tables over bullet points, figures in thousands. These rarely go stale and cost nothing when wrong.
• Who you are. "I'm an analyst in the FP&A team; my audience is the CFO" saves a paragraph of context per chat.
• Recurring vocabulary. Product names, team names, the acronyms nobody outside your company knows.

Where it quietly hurts

• Definitions that change. Fiscal calendars, pricing tiers, which customers count as "enterprise". A wrong memory produces a confident answer with no visible assumption in it.
• Clients and contexts bleeding together. An agency analyst working for two retailers doesn't want one's margin targets turning up in the other's analysis.
• Invisible inputs. When you share Claude's answer, your reader can't see the memory it used. The answer looks self-contained and isn't.

For business definitions, we'd use a project's instructions instead (see Claude Projects). They are visible, scoped and dated in a way memory isn't. The honest counter-point: memory needs no upkeep at all, and for one person working alone on one company, that convenience often wins.

Nobody has published an accuracy study of Claude's memory, and Anthropic's help page carries no warning that memories can go out of date. They can, obviously. The risk is less that Claude misremembers and more that it remembers correctly something that has stopped being true.

There's a broader reason not to load every conversation with more context. Anthropic's own engineering team describes "context rot": as the context grows, the model's recall of any one piece of it falls (Anthropic Engineering, September 2025). Memory adds context to every chat whether the question needs it or not.

A two-minute check

1. Read your topics. Open Settings > Memory and skim the list. Anything about prices, targets, team structure or definitions older than a quarter deserves a second look.
2. Ask about the answer in front of you: "What did you assume about me or my company in that answer, and where did each assumption come from?"
3. Run it incognito. Paste the same question into an incognito chat. If the answers differ, memory changed the result, and you now know to find out how.
4. Fix the source. Edit or delete the topic that caused the difference, rather than correcting Claude in the chat and hoping.

The general version of step 2 is what a reasoning chain makes routine. Count maps each claim in a Claude answer back through the prompt, what the model already believed, and each search or calculation, and labels claims that rest on something nobody checked as assumptions. A stale memory is that kind of something.

Memory makes Claude feel like a colleague who has been around a while. Like that colleague, it is sometimes working from how things used to be. ChatGPT's version works differently.

Does Claude have memory?

How do I make Claude forget something?

If I delete a chat, does Claude forget it?

Does Claude's memory carry across projects?

How do I stop Claude remembering a conversation?