What Claude Code remembers
Anthropic's documentation (How Claude remembers your project) describes two ways knowledge carries from one session to the next:
- CLAUDE.md files are instructions you write: how to build the project, rules to follow, things to avoid. They can sit in the project, in your user folder or be set by your organisation, and every one that applies is loaded at the start of each session.
- Auto memory is notes Claude Code writes for itself as you work: your corrections, preferences and what it has learned about the project. When you say "remember that the tests need a local database", this is where it goes.
Both are plain text files you can open and edit. In a session, /memory lists
them, opens them, and turns auto memory on or off.
What loads at the start of each session
Auto memory lives in a folder for each project,
~/.claude/projects/<project>/memory/, with a file called MEMORY.md as its
index. Only the first 200 lines or 25KB of that index, whichever comes first, are loaded when
a session starts. The detailed notes sit in separate topic files that Claude Code opens when
it needs them.
We build Zaaheen with Claude Code every day, so we measured our own project's memory after five months of sessions:
| Option | Kilobytes |
|---|---|
| All memory files | 494 KB |
| Loaded at start | 19 KB |
| Start-up limit | 25 KB |
So Claude Code does remember, but most of what it knows is looked up on demand rather than read every time. Keeping the index short, with one line per memory, is what keeps the important things in view. Anthropic suggests keeping each CLAUDE.md under 200 lines for the same reason.
What it doesn't carry
The documentation is clear about the edges:
- One memory folder per project. All the worktrees of one repository share it, but a different repository has its own, so what Claude Code learned about you in one project isn't in the next.
- One computer. Auto memory is machine-local and isn't shared across machines or cloud environments.
- One app. Cursor, Codex, ChatGPT and Claude Desktop don't read Claude Code's memory folder.
For project rules, that is exactly right: they belong with the project. For facts about you, such as how you like answers, the tools you prefer or what you're working on this month, it means telling each project and each app again.
Remembering you across projects and apps
A memory that lives outside Claude Code fills that gap. Claude Code can connect to tools through MCP, the standard way AI apps plug into tools and data, so it can save facts about you to a memory that every project, and your other AI apps, can read.
Zaaheen is one such memory, for Windows and Mac. You connect Claude Code once, with one command from Zaaheen's Agents tab, and it is then available in every Claude Code project (step by step). Your memories stay encrypted on your own computer.
A good split: keep project rules in CLAUDE.md, let auto memory hold what Claude Code learns about the project, and keep what's true about you in one shared memory that every app reads.