Claude Project Memory: What Carries Over Between Chats

Claude project memory is the memory Claude keeps for one project: a separate memory space and project summary, built from the chats in that project and kept apart from your other projects and from chats outside any project. When memory is on, a new chat in the project starts already knowing what earlier chats there established — who the client is, which format you prefer, that the deadline moved. What does not carry over is the conversations themselves: earlier chats are not loaded into the new one, and anything that must be exact belongs in the project’s knowledge, not in memory.

What Claude’s memory is

Anthropic’s article on chat search and memory↗ (opens in a new tab) describes memory as Claude saving context as a set of individual topics while you chat, rather than summarising conversations after they end. It saves on its own, and you can also say “remember this”. The article lists the kind of thing it keeps:

• Your role, projects and professional context.
• The people and places in your work and life.
• Communication preferences and working style.
• Technical preferences and coding style.
• Project details and ongoing work.

Sensitive subjects — health, religious beliefs, politics and similar — are left out by default; you can opt in under Settings, then Memory. Some things are never saved even if you ask: government ID numbers, criminal history, financial account numbers and immigration status.

Memory is on by default on Free, Pro and Max plans on the web, desktop and mobile apps. On Team and Enterprise plans it is off until an owner turns it on for the organisation, and it is not available to organisations with HIPAA, public-sector or custom data retention agreements.

How project memory differs from memory outside projects

The same article is specific: each project has its own separate memory space and dedicated project summary, so the context in each project stays focused and separate from other projects and from non-project chats. A preference you set while drafting board papers in one project does not leak into a chat about your holiday, or into another client’s project.

That makes where a chat lives matter. The help page on creating and managing projects↗ (opens in a new tab) uses exactly this example: if you start an unrelated chat inside a project by mistake, choose “Remove from project” from the dropdown next to the chat name, and it is included in your non-project memory instead. Moving chats in and out of projects is how you manage which memory they feed.

What a new chat in a project starts with

• Project knowledge: every file you uploaded, in every chat. On paid plans a large knowledge base is searched rather than loaded whole.
• Project instructions: applied to every chat in the project.
• Project memory: what Claude has saved from earlier chats in this project, when memory is on.
• Chat search, on paid plans: Claude can search earlier conversations when you ask — “what did we decide about the pricing page?” — but only within this project. It shows as a tool call, with citations back to the original chats.
• Not the other chats themselves. The help centre says context is not shared across chats within a project unless the information is added to the project knowledge.

The practical rule follows from that list. Memory is Claude’s own summary, good for preferences and background. A decision, a figure or a definition you will need word for word should be written into the project’s knowledge, where you control the wording. Claude Projects best practices has a simple way to do it: end the chat by asking for a short write-up and add it to the knowledge.

Viewing, editing and turning it off

• See it: Settings, then Memory. Everything Claude remembers is listed under Topics; open one to read it, use the edit icon to change it, or Delete to remove it. A fix applies to every conversation from then on.
• Change it from a chat: tell Claude what to remember, change or forget, and it applies to your next conversation.
• Stop searching old chats: switch off “Search and reference chats” in the same settings. It is separate from memory.
• Pause memory: Claude keeps what it has but neither uses it nor saves anything new. Chats during the pause are not added later.
• Reset memory: permanently deletes all memories, including project memories. It cannot be undone.

One detail is easy to miss. Deleting a chat does not remove the memory entries that came from it; the help centre says you delete those individually. If a chat held something you want gone, delete the chat and then check Topics.

The help centre also notes that a small number of Team and Enterprise organisations still have the older memory experience, which lives under Settings, then Capabilities, and builds a summary that updates every 24 hours. If you see Memory directly in Settings, you have the current one described here.

Incognito chats, and why they sit outside projects

An incognito chat is Claude’s way to leave no trace in memory. Anthropic’s page on incognito chats↗ (opens in a new tab) says they are available on every plan, are not saved to your chat history, do not use your existing memory, and are not added to it later. They are started from the ghost icon when you open a new chat.

• Incognito is only available outside projects. There is no ghost icon inside a project, so a sensitive one-off question belongs in a separate incognito chat, not in the project.
• Your profile preferences still apply in an incognito chat.
• An incognito chat cannot be converted to a normal chat or reopened once closed. Copy out anything you need first.
• It is not invisible. Incognito chats are retained for 30 days by default, and on Team and Enterprise plans they are included in organisational data exports.

Memory in Cowork, and a different product with the same word

When Claude Cowork runs in the cloud, memory is shared between chat and Cowork in both directions; local Cowork sessions do not use memory. A Cowork project’s memory is scoped to that project, like a chat project’s. The surface-by-surface detail is in Claude Projects in chat vs Cowork.

Claude Code is separate again. Its memory is CLAUDE.md files and auto memory on your machine, covered in agent memory for Claude Code. The newer projects in Claude Code keep project memory as files, with a MEMORY.md index that every thread reads, according to Anthropic’s Claude Code projects documentation↗ (opens in a new tab). None of these share a store with the chat app’s project memory.

What memory cannot do for a team

• It is personal. On Team and Enterprise plans each person manages their own memory, and the help centre says owners cannot view or edit a user’s individual memories. A colleague in the same shared project has a different memory of it.
• It is not a log. The help centre says individual memory edits are not recorded in audit logs, so there is no answer to “when did Claude start assuming that?”
• Turning it off at the organisation level deletes everyone’s memory at once, permanently.

For what a whole team, and every assistant it uses, should know, a written record works better than any one person’s memory. On fenbs that is the board’s AI context: notes every connected assistant reads before it starts, each signed with who wrote it, beside tasks whose every change is recorded. The general case — short-term, long-term and shared memory — is in what is agent memory.