A project in the Claude apps (claude.ai, desktop, mobile) is a workspace for one area of work: a client, a codebase, a side project. It gives every chat in it the same background, so you can start a fresh chat for each task without re-explaining everything. This is the chat-side counterpart of what CLAUDE.md and auto memory do in Claude Code.
How it works
A project has three layers:
- Instructions and files. What you put there by hand: project instructions and the project knowledge base (documents, code, other files). It is static and used across the chats of the project. When the knowledge gets close to the context limit, Claude switches to retrieval (RAG): it searches the files for the relevant parts instead of loading all of them.
- Project memory. What Claude remembers on its own from the chats in this project: your role and work, decisions, how you like answers. In the current version, Claude saves memory as separate topics while you chat, not as a summary after the chat ends, and you can see them listed under Topics.
- Past-chat search. On paid plans (Pro, Max, Team, Enterprise), Claude can search earlier conversations. Inside a project, the search is limited to that project’s chats.
Each project has its own memory. Claude combines a project’s memory with your account memory from non-project chats, but never with the memory of other projects. What you discussed in a “flat rental” project does not leak into a “work” project, and the other way round.
Why it matters
Memory is what makes “one task, one chat” practical. Instead of dragging one huge chat through a week, which degrades as its context window fills up, you start a new chat per task: the context is fresh, and the important bits from earlier chats come back through memory and search. See One task per session.
What to do about it
- Steer it from the chat. Tell Claude what to remember, change or forget; the update applies to your next conversation. You can also open a topic to read, edit or delete it.
- Put critical rules in the project instructions. Memory is Claude’s own condensed notes, so it is the less reliable place for something that must always hold.
- Turn memory off when you do not want it. In a chat with memory off, Claude does not use what it remembers and saves nothing from that chat to memory; the chat still stays in your history. An incognito chat is not saved to memory or history at all, and search skips it.
Gotchas
- A summary, not a transcript. Memory keeps topics, not the full conversation, so details can get lost and facts can go stale.
- Plans differ. Memory is on by default on Free, Pro and Max. On Team and Enterprise the owner decides whether it is available, and each member has to turn it on. Past-chat search needs a paid plan.
- Project knowledge costs context. Files in the project feed every chat, which is why big knowledge bases fall back to retrieval.
Related
- Auto memory: the same idea in Claude Code.
- Context window: why fresh chats beat one long one.
- One task per session: the habit that memory makes painless.
- Coding agent vs chat assistant: how memory differs between the two.