Sign in

Shared memory for Claude Code across computers and people

Guides · Updated

Claude Code remembers through CLAUDE.md files and its own auto memory, and both live on one computer. To share what it learns across your computers or with a teammate, you either commit memory files to a repo or give every agent one shared place to read and write.

How Claude Code memory works today

  • CLAUDE.md files. Instructions Claude Code loads at the start of a session: your user file (~/.claude/CLAUDE.md), the project’s CLAUDE.md, and a personal CLAUDE.local.md you keep out of Git. A file can pull in others with @path imports.
  • Rules files. Instructions can be split into files under .claude/rules/. A rule can be limited to certain paths, so it only loads when Claude works on matching files.
  • AGENTS.md. Claude Code reads a project’s AGENTS.md when there is no CLAUDE.md. To share one file with other coding tools, the docs suggest importing it with @AGENTS.md from CLAUDE.md.
  • Auto memory. Notes Claude Code writes for itself as it works, kept under ~/.claude/projects/ with a MEMORY.md index that loads each session.
  • The /memory command shows these files, opens them for editing and turns auto memory on or off.

Where it falls short

  • Auto memory stays on the computer that wrote it. Your laptop’s agent never sees what your desktop’s agent figured out.
  • A project CLAUDE.md can be committed and shared, but someone has to write and merge every change by hand.
  • Your co-founder’s agent repeats the mistake yours already learned from.

Your options for sharing it

  • Commit it. Put shared rules in the project’s CLAUDE.md, or in one file that each CLAUDE.md pulls in with an @ import, and keep it in Git. Free and transparent, but only for that repo, and every lesson waits for a person to write it down and merge it.
  • Sync the memory folder. Auto memory is plain Markdown files on disk, so you can sync that folder between your own computers with Git or a file-sync tool. It works for one person; two people’s agents writing to the same files will clash.
  • Run a memory MCP server. MCP servers such as the reference knowledge-graph memory server give an agent a store it can add to and search. You host it, and it is up to you to keep it tidy and to share it safely.
  • Use a shared Brain. pairpipe hosts one memory for a pair, across both people’s computers, and keeps it tidy for you. The rest of this page explains how.

A Brain the whole pair shares

Each pair gets a Brain: notes that every connected agent can search and write, on any computer, for both people. When an agent learns something worth keeping (a fix, a rule, a preference), it first looks for a similar note, then saves the lesson as a short note: when it applies, the rule, and why. The next session, anywhere in the pair, can find it.

What loads at the start of a session

Not everything. A small, fixed budget is filled with the pair’s Shared rules note (a shared AGENTS.md) and the strongest lessons. It is framed as information, not commands, and each person’s trust setting decides what their own agent loads. Everything else stays searchable on demand, so the context window stays small.

Keeping it tidy without you

  • A lesson close to an existing one strengthens it instead of adding a duplicate.
  • A lesson that contradicts another is kept and marked as a conflict, so you can see both.
  • Lessons nobody uses fade over time and retire on their own.
  • You can read what your agents learned each day. From that feed you can undo a lesson being learned, merged, retired or flagged as a conflict, an approval or discard, or a change to the Shared rules.

It works alongside CLAUDE.md rather than replacing it. Keep repo-specific instructions in the repo, and let the Brain carry what should follow you and your partner everywhere. To start, sign in, press Show my setup line and paste it into Claude Code. Working alone, pair two of your own computers. Codex works too, in beta. Steps are in Help.

More guides