Use memory from an external agent

Search, read and save Agent Fleet memory from Claude Code or Claude Desktop on your own computer, or from claude.ai. You can also bring the memory your own Claude Code kept into Agent Fleet memory.

How it works

What you do	What the agent does
Allow memory to be used from outside	—
Connect your client to /mcp or /mcp/memory	Receives the Agent Fleet memory tools
Ask it to “search the memory” or “save this to the memory”	Searches and saves memory in the project of the repository it works in

A memory saved from outside is read by every kind of session in the workspace. It shows in the memory Changes view, where you can revert it.

Before you start

1. In ⚙ Settings → Agent memory → Settings, choose “Use Agent Fleet memory from outside Agent Fleet”. It is Off by default

   

   Choice	What an outside client can do
   Off	Nothing
   Read only	List, search and read memories
   Read and write	Read, and also save, patch, report and forget

2. Keep the workspace running. The memory lives inside it, so it cannot be used while the workspace is stopped

If your tenant administrator has blocked memory from outside, the memory tools never appear, whatever this setting and your token say (for administrators).

Connect a client

Where you connect depends on the client.

Client	Connect to	How
Claude Code or Claude Desktop on your computer	/mcp (session tools too) or /mcp/memory (memory only)	Issue a token
claude.ai	/mcp/memory	Connect by signing in

The setting’s “How to connect” opens MCP tokens & apps, which lists what a token can do and links back to this setting.

• The narrower of the token’s scope and the setting wins. Reading needs read or higher; saving and forgetting need both write and “Read and write”
• A token or connection made for /mcp/memory does not work on /mcp, and the other way round. To switch, connect again

Use /mcp/memory for claude.ai

claude.ai hands every connector you add there to all the Claude Code sessions of your account, including the sessions in this workspace.

• A connector to /mcp would give those sessions your full member authority (Do not connect this deployment’s /mcp to a session)
• A connector to /mcp/memory serves the eight memory tools and nothing else. The workspace’s sessions already use the same memory, so it gives them nothing new
• The consent page says the app is asking for your memory only

A claude.ai connector you already have on /mcp keeps working. To move it, remove it in claude.ai and add it again with the /mcp/memory address. Keep /mcp for a client that runs only on your own computer and needs the session tools, such as Claude Desktop or a command-line client.

Ask the agent

Just ask your client in plain words.

Search my Agent Fleet memory for how we run the integration tests

Save to Agent Fleet memory that this repo’s lint step needs make gen first

Tool	What it does	Needs
memory_projects	Lists the projects and their repository addresses	Read
memory_index, memory_search, memory_read	Show the index of memories, search, read	Read
memory_save, memory_patch	Save a memory, change part of one	Read and write
memory_feedback	Report a memory that failed you	Read and write
memory_forget	Forget a memory	Read and write

• The project follows the repository being worked in. The agent passes the output of git remote get-url origin, and Agent Fleet matches it against each project’s repository. https://, ssh:// and git@ spellings all work
• Saving and forgetting need a scope. Use project for something about one repository, and user only for something that holds in every project. A call without one is refused, so a fact about one repository cannot spread to every project by mistake
• Write facts only. Text that looks like a secret is refused. A save goes through the same revision check, secret scan and limits as an agent’s save inside the workspace

Bring your own Claude Code’s memory in

Import from Claude Code, under Import & export, reads only the Claude Code inside the workspace. The memory of Claude Code on your own computer comes across by having that Claude save each memory with memory_save.

Prepare

1. In ⚙ Settings → Agent memory → Settings, turn on “Let sessions use Agent Fleet memory”, so the workspace’s sessions read what you bring in
2. On the same screen, set “Use Agent Fleet memory from outside Agent Fleet” to “Read and write”
3. Connect Claude Code on your computer to /mcp with the write scope, by signing in or with a token (commands)
4. Clone the repository under ~/repos in the workspace. A clone that holds no memory yet is started as a project by the first save

Save the skill

Save this skill once on your computer, at ~/.claude/skills/af-memory-import/SKILL.md.

---
name: af-memory-import
description: Copy this project's Claude Code memory into Agent Fleet memory
disable-model-invocation: true
---
Copy the memory you keep for this project into Agent Fleet memory. Change nothing in your own memory files.

1. Run `git remote get-url origin`; that is `repository` for every call below. Call agent-fleet
   `memory_projects` and check that it lists this repository (a repository cloned in the workspace is listed even before it holds any
   memory, with a count of 0). If it does not, the repository is not cloned in the workspace: stop and tell me.
2. List the files in your memory directory, except the index `MEMORY.md`. For each file, show me its name,
   type and description in one table, and propose a scope: `project` for anything about this repository,
   `user` only for a fact that holds in every project. Wait for my OK on the table.
3. For each file I approved:
   - Call `memory_search` with its name. If a memory of that name exists, `memory_read` it; when the text is
     the same, skip it; when it differs, show me both and ask which to keep. To replace it, pass the
     `revision` you read.
   - Call `memory_save` with `scope`, `repository`, `name` (lower case; letters, digits, `-`, `_`, `.`;
     at most 64 characters), `description` (one line), `type` (`user`, `feedback`, `project` or `reference`)
     and `body` (the file's text without its frontmatter).
   - If a save is refused, for example as a possible secret, do not retry it in other words. Note the reason.
4. Report what was saved, skipped and refused.

Bring it in

1. In a Claude Code session of that repository on your computer, type /af-memory-import
2. Check the table Claude shows you and give your OK
3. When it has finished, look at the memory Changes view. The saves are listed with author kind external, and each can be reverted

• Your own memory files are not changed. What your own Claude Code learns later stays there, so run the skill again when you need it. A memory with the same text is skipped
• Instruction files (CLAUDE.md) are not memory. Put what they say in ⚙ Settings → Agent instructions, or keep the repository’s own file in the repository

What is recorded and what leaves

What	What happens
A memory saved from outside	Listed in the memories with author kind external. The commit carries the name the client gave for itself, for information only
Each save, patch, report and forget from outside	Written to the audit log with the project id. The memory’s name is not
The memory text	Reaches that client’s model, outside the workspace’s egress controls

The secret scan stops credentials, not customer names or internal context. Think about whether the content may leave before you allow it.

When it goes wrong

The memory tools do not appear

Check these in order.

1. “Use Agent Fleet memory from outside Agent Fleet” is not Off
2. The token’s or connection’s scope is enough (write to save)
3. Your tenant administrator has not blocked it. When they have, no tool appears, whatever the setting and the scope say. Ask them if you are not sure

The answer says to start the workspace

The workspace is stopped. Start it in the Console and ask again.

A save is refused because the repository is not found

Situation	What to do
It is not cloned in the workspace	Clone it under ~/repos and save again
The workspace has two or more clones of it	Pass an id from memory_projects as project
It is on a Git server you run yourself	Only github.com, gitlab.com, bitbucket.org and codeberg.org are matched. Pass project

A save is refused asking for a scope

A save or forget without a scope is refused. Ask the agent to save again with project for something about the repository, or user for something about every project.

Related

• Drive your workspace from an external Claude: issue a token and register it in your own Claude Code
• Connect a remote MCP client by signing in: connect claude.ai and others without a token
• Import and export memory: import the memory of the workspace’s Claude Code or muse