What’s the best memory for Claude Code?

by Benjamin a. Mowinckel · checked 2 November 2026

Start with what’s built in: CLAUDE.md, the instructions you write, and auto memory, the notes Claude Code writes for itself. Add claude-mem, Basic Memory, an Obsidian vault, or a record you own, such as alexandria., only when you need memory to search, to cross projects, or to reach your other tools.

CLAUDE.md and auto memory suit almost everyone, and need nothing installed. Claude Code reads your CLAUDE.md files at the start of every session, one for you across all projects and one per project, and it can read an AGENTS.md instead. Auto memory is on by default and keeps its notes in plain Markdown, one folder per repository, loading the first 200 lines or 25KB of its index each session. What it learns in one repository stays in that one.

claude-mem suits someone who wants every session recorded and searchable. It’s an open-source Claude Code plugin that records what the agent does, summarises it, and brings the relevant parts into later sessions, kept in a database on your computer, with an optional hosted service.

Basic Memory suits someone who wants memory they can read. It’s an open-source MCP server, the standard Claude Code uses to reach outside tools, that keeps notes as Markdown files on your disk, which Obsidian can open as a vault. It has a paid hosted version too.

An Obsidian vault suits someone who already writes there. Run Claude Code in the vault, or point it at the vault from your CLAUDE.md, and your notes become its memory with nothing extra installed.

A record you own suits someone who uses more than Claude Code. That’s what I built alexandria. to be, a blueprint your own model reads and builds on your computer: one record of how you think, which Claude Code, Codex, and Cursor read before they help and save back to, across every project.

Compared on each tool’s own documentation, 2 November 2026.

What does alexandria. add to Claude Code’s own memory?

It leaves CLAUDE.md and auto memory exactly as they are. Your model writes one guide, the AGENTS.md at the top of your record, that every tool reads first, and Claude Code’s hooks load it at the start of each session and copy each conversation into your folder. Where Claude Code keeps its memory on your computer, your model reads it there and draws what it learned about you into your record, so Codex and Cursor know it too.

Questions

Is auto memory enough on its own?

For one project, often yes. It is scoped to each repository, so what Claude Code learns about you in one project doesn’t carry to the next, which is where a memory that crosses projects starts to pay.

Will adding memory slow every session down?

Anything loaded at the start of a session uses room in the context window. That’s why auto memory loads only the first 200 lines or 25KB of its index, and why a good record points to files to read when needed rather than loading everything.

Can I turn auto memory off?

Yes, from the /memory command in a session, or for one project in that project’s settings.

Does this work on Windows?

Yes. Claude Code runs on Windows, and so do the tools above.

Can I use it on a work laptop?

That depends on what your employer allows you to install. With alexandria., you can tell your model to leave work files out, now or any time later.

What if I use Claude Code in the cloud?

A cloud session keeps no record on your own computer. Work committed to the repository stays, while files saved only in the session can disappear, so memory that lasts needs a tool on your computer.

What does leaving cost?

Nothing, for the built-in memory, Basic Memory, a vault, or alexandria., because each is plain files. One command takes alexandria. out of your tools and leaves your files where they are.

Show the blueprint to Claude Code, and it’ll tell you what you’re missing, if anything.