Claude Memory: What It Remembers, How to Control It, and What It Won't Do

On March 2, 2026, Anthropic turned memory on for every Claude account, free plans included, and launched a tool the same day to import your memories out of ChatGPT. The second part tells you what the first part was for.

Before any of the how-to, one piece of housekeeping that most guides skip. Three unrelated Anthropic features are called memory, and searching the term drops you into whichever one Google decides you meant. This guide covers the chat feature: the thing that remembers you between conversations at claude.ai and in the apps. The other two get sorted out below.

What is Claude memory?

Claude memory is a feature that lets Claude retain facts about you across separate conversations, so it stops starting from zero every time you open a new chat. Anthropic's documentation describes what it focuses on: your role, projects, and professional context, your communication preferences and working style, your technical and coding preferences, and the details of work you have in progress.

The word doing the work there is retain, not record. Claude isn't keeping your conversations. It writes short summarised entries about you, which is why memory can hold "prefers concise answers, works in higher education" without storing the chat where you said it.

The stack is your conversations. The card on top is what memory actually keeps.

Which Claude memory are you looking for?

Three different Anthropic features share the name, and they have almost nothing to do with each other.

If you type "claude memory" into a search box, the results mix all three freely, which is how people end up reading about markdown files in a repository when they wanted to stop Claude forgetting their job title. Everything below is the first row.

The other two are worth one line each so you can rule them out. Claude Code's memory is a pair of systems for coding work: CLAUDE.md files you write by hand, and an auto memory where Claude saves its own notes about a codebase, kept in a project folder and loaded at the start of each session. The memory tool is an API feature for developers building applications, where Claude asks your app to read and write files in a store you host yourself.

How does Claude decide what to remember?

Claude synthesises memory rather than recording it, reviewing recent conversations roughly every 24 hours and pulling out facts, preferences, and patterns that look worth keeping. Those become short structured entries you can read.

That 24-hour cycle explains something that confuses people: you tell Claude something on Monday, check your memory settings, and find nothing there. Nothing is broken. The synthesis simply hasn't run.

There's a faster path, and it's the one worth using. Say "remember that I work in higher education" or "add to memory that I prefer short answers" in any conversation, and Claude writes the entry immediately rather than waiting for the next cycle. Automatic memory is passive and slow. Telling it directly is instant and precise, and in practice a memory profile you shaped deliberately beats one assembled from whatever you happened to mention.

Where do you turn it on and manage it?

Everything lives in one place: open Settings, then Memory. Two toggles run the feature. "Generate memory from chats" is the one that builds and uses your memory, and "Search and reference chats" lets Claude look back through past conversations. They're independent, which matters more than it sounds and comes up again below.

The same screen shows what Claude has stored, organised by category, and every entry can be edited or deleted individually. Worth doing occasionally: memory built by synthesis picks up the odd wrong inference, and a stale entry about a job you left quietly shapes answers for months.

One exception sits outside all of it. Incognito chats, marked with a ghost icon, are temporary and unsaved, and Claude neither remembers nor searches them. That's the switch for a one-off conversation you don't want influencing anything later.

Pause or reset: two controls that do different things

Pause keeps everything Claude has learned about you but stops it adding anything new. Reset permanently deletes the lot. People reach for the wrong one regularly, because both sound like "turn memory off."

Pause is the reversible, everyday option. If you're doing a stretch of work that isn't representative, researching a topic you'd rather not be permanently associated with, or covering for a colleague in a field you don't normally touch, pausing keeps your existing profile intact while that work stays out of it.

Reset is the nuclear one and there's no undo. Anthropic notes that memory data is included in your data exports and retained under the same policies as your chat history, so if there's anything in there worth keeping, export before you clear.

What Claude memory doesn't do

Four limits are worth knowing, and none of them are flaws exactly, just the shape of the thing.

It's reactive. Memory builds from what you happen to discuss, starting empty and filling in slowly, so a brand-new account has nothing and a heavily used one has a lot. Nothing is prefilled.

It stays inside Claude. Your memory lives on Anthropic's servers and works in Claude alone. Move to ChatGPT, Gemini, or a coding tool and none of it travels with you, which is the reason the import tool exists in the first place.

It isn't shareable. Memory is personal to your account, with no built-in way for a team to build a shared layer, so five colleagues using Claude for the same project each teach it the same context separately.

And it's shallow by design. Memory holds a summary of who you are and what you prefer, not the reasoning behind decisions you made or the artifacts you produced. It knows you like short answers. It doesn't know why you rejected the second option in a plan three weeks ago. For work you want to persist properly, Claude Projects holds documents in a knowledge base, and Claude Artifacts keeps the things you actually built.

The free and paid split that catches people out

Memory generation is free and chat search is not, which is a genuinely odd combination once you notice it. Anthropic's documentation lists memory across free, Pro, and Max plans on web, desktop, and mobile, with Enterprise and Team on their own rollout schedules. Cowork doesn't have it at all.

Searching and referencing your past chats, though, is a paid feature. So a free user gets a Claude that quietly learns their preferences and applies them, but can't ask it to find that conversation from last Tuesday. The synthesis works; the retrieval doesn't.

That distinction is easy to miss because both toggles sit in the same settings panel, one above the other, with nothing marking one as paid. If you're on a free plan and the second toggle seems to do nothing, that's why.

How memory behaves inside Projects

Each project keeps its own memory space and its own project summary, separate from your other projects and from your ordinary chats. Your global memory profile doesn't reach inside a project, and what Claude learns in a project stays there.

The separation is deliberate rather than a limitation, and it's the right call: work context and personal context genuinely shouldn't bleed into each other. The catch is that it surprises people. Something Claude knows about you in a normal chat won't be there when you open a project, and you'll teach it twice. Our guide to Claude Projects covers how that isolation works alongside the knowledge base.

Can you bring your memory from ChatGPT?

Yes, and the mechanism is more low-tech than you'd guess. Anthropic launched an import tool at claude.com/import-memory on March 2, 2026, the same day memory went free, covering ChatGPT, Gemini, and other assistants.

It doesn't connect to the other service. Anthropic hands you a prepared prompt that you paste into ChatGPT or Gemini, asking that assistant to write out everything it has stored about you in a format Claude can read, and you bring the result across. No API, no account linking, just one chatbot dictating its notes to another.

Two things follow from that. It's a migration and not a sync, so anything you teach ChatGPT next month stays in ChatGPT. And the quality of what arrives depends entirely on how well the other assistant summarises itself, which is worth checking rather than trusting. If you're weighing the two assistants more broadly, our ChatGPT vs Claude comparison covers the wider picture, and ChatGPT's nearest equivalent to a standing memory profile is covered in ChatGPT custom instructions.

Releasing a defection tool on the same day you make memory free is not subtle, and it's aimed squarely at the thing that keeps people on an assistant they've outgrown: the months of context they'd have to rebuild.

What this post does not cover

This is a guide to the consumer memory feature in Claude, not a manual for Claude Code's CLAUDE.md system or the developer memory tool, both of which work differently and are aimed at other jobs. It isn't privacy or compliance advice, so check your own organisation's policy before letting an assistant retain work context. Plans, toggles, and which tiers get which piece change often, and Anthropic's own help pages are the place to confirm current details before relying on them.

Sources

1. Claude Help Center: use Claude's chat search and memory to build on previous context
2. Claude Code docs: how Claude remembers your project
3. Anthropic API docs: the memory tool
4. MacRumors: Anthropic adds free memory feature and import tool
5. Fast Company: Claude can now take your memories from ChatGPT, Gemini, and Copilot