Get Harbor →

Blog

Claude Code team memory: what is shared, what is not

Published: October 28, 2026

Claude Code has no team memory you can turn on. What your team shares today is what you commit to git, what an admin pushes to every machine, and skills. Auto memory, the notes Claude writes for itself, stays on one person's machine. Below is every place Claude Code reads memory from, who each one reaches, and how teams close the gap.

Does Claude Code have team memory?

Not as a feature you can use, as of 24 October 2026. The feature request for shared team memory has been open since 25 March 2026, with 29 comments and no answer from Anthropic. Its first line sums up the state of things: “Claude Code's memory system is individual-only.”

There is a trace of something on the way. The Claude Code changelog for v2.1.172 (10 June 2026) lists a fix for “memory recall not finding mounted team memory stores” in remote sessions, and a bug report from 5 June says “Team memories don't actually work.” No docs page describes the feature, so do not plan around it yet.

Which Claude Code memory files does my team share?

From Anthropic's memory docs, in the order Claude Code loads them, broadest first:

MemoryWhere it livesWho it reaches
Managed policy CLAUDE.md/Library/Application Support/ClaudeCode/CLAUDE.md (macOS), /etc/claude-code/CLAUDE.md (Linux, WSL), C:\Program Files\ClaudeCode\CLAUDE.md (Windows)Everyone in the organization, pushed by IT
User CLAUDE.md~/.claude/CLAUDE.md and ~/.claude/rules/Only you, in every project
Project CLAUDE.md./CLAUDE.md or ./.claude/CLAUDE.mdThe team, through git
Project rules.claude/rules/The team, through git
Local CLAUDE.mdCLAUDE.local.md, kept out of gitOnly you, in this project
Auto memory~/.claude/projects/<project>/memory/Only you, on this machine

So the shared layer is whatever lives in the repo, plus whatever an admin installs. Everything Claude learns on its own lands in the last row.

Is auto memory shared with my teammates?

No. The docs are plain: “Auto memory is machine-local” and its files “are not shared across machines or cloud environments.” It is on by default. At the start of a session Claude Code loads the first 200 lines or 25KB of its MEMORY.md, whichever comes first. You can switch it off with the /memory toggle, autoMemoryEnabled: false in settings, or CLAUDE_CODE_DISABLE_AUTO_MEMORY=1. The rest of how it works, for one person, is in Claude Code auto memory.

That is why a teammate's Claude makes the mistake yours was corrected on last week: the correction went into your auto memory, or nowhere. Our note on how to make Claude Code remember your corrections covers turning one into a rule.

Can Claude's own notes be committed to git?

For subagents, yes. A subagent with memory: project in its frontmatter writes to .claude/agent-memory/<name>/, which the subagent docs call “shareable via version control”. It is the one documented memory that Claude writes and your team can commit. The main session's auto memory has no project scope.

How do admins push one CLAUDE.md to every developer?

  • On the machines. Put a managed policy CLAUDE.md at the paths in the table, with MDM, Group Policy or Ansible. Users cannot exclude it with claudeMdExcludes.
  • From claude.ai. On Team and Enterprise plans, an owner sets a claudeMd under Organization settings, Claude Code, Managed settings. It loads before user and project files. Sessions signed in with a Console API key do not get it (server-managed settings).

In what order do CLAUDE.md files load, and which one wins?

None of them wins. “All discovered files are concatenated into context rather than overriding each other”: managed first, then user, then project, then CLAUDE.local.md, appended after the project file. Files in subfolders load when Claude works there. The whole set arrives as a user message after the system prompt, and the docs say Claude treats it as “context, not enforced configuration”. They suggest keeping each CLAUDE.md under 200 lines.

Three things you will still read elsewhere are out of date. The September 2025 docs called CLAUDE.local.md deprecated, allowed imports five hops deep, and said files higher in the tree take precedence. Today CLAUDE.local.md is supported, @ imports stop at four hops, and nothing overrides anything.

How do we share skills and plugins?

Commit a skill under .claude/skills/ and everyone in the repo has it, or ship it in a plugin or managed settings to reach the whole team (skills docs). A plugin enabled in the committed .claude/settings.json is offered to teammates, but not downloaded for them. Skills your organization adds on claude.ai sync into Claude Code for anyone signed in with a claude.ai account, from v2.1.273.

What does Claude Code still not share?

  • What a session learned. It goes to that person's auto memory, on that machine.
  • A correction made in chat. It ends with the session unless someone writes it down.
  • Anything decided outside the repo: a Slack thread, a review comment, a call. A file only knows what somebody copied into it.
  • Your teammates' other agents. Cursor, Codex and Copilot do not read Claude's auto memory, and each keeps its own (see Claude Code memory vs Codex, Cursor and Copilot).

How teams fill the gap

  • Commit more. Write corrections into CLAUDE.md or .claude/rules/ in the pull request that needed them. It works, and it depends on someone remembering to.
  • Commit agent memory. Give the subagents your team relies on memory: project, and review what they write like any other diff.
  • Push from the top. Use managed settings for the rules every repo shares.
  • Use a shared memory. An MCP memory server or a company brain that every agent reads. Our list of MCP memory servers for teams compares them.

Harbor is the last kind: shared memory for AI coding agents. It learns from where your team decides (Slack, pull request reviews, docs and agent sessions), lets you set what waits for a person, and gives Claude Code, Cursor, Codex and GitHub Copilot the rules that apply to each task. To use the same rules in every agent without it, see how to share memory and rules across tools.