Set how memory is used

Agent memory is off by default. Turn on one switch and every kind of session reads and writes the same memory. What the switch does to the agent’s own memory is summed up in Choose which memory your agents use.

How it works

What you do	What the agent does
Turn on “Let sessions use Agent Fleet memory”	From the next session it starts, reads and writes memory with af’s memory_* tools
Turn it off	Reads and writes are refused, in running sessions too. Saved memories stay
Choose the outside, tidy-up and lint switches	Each of those runs only while its switch is on

Turn the switches on

Open ⚙ Settings → Agent memory → Settings. This view works even while the workspace is stopped; the switches are saved for the next start.

Switch	What it decides	Default
Let sessions use Agent Fleet memory	Whether sessions get the memory tools	Off
Use Agent Fleet memory from outside Agent Fleet	Whether agents on other machines and connected apps may use it (Off, Read only, Read and write)	Off
Memory tidy-up (suggestions)	Whether a model drafts fixes (Use memory tidy-up)	Off
Automatic lint marks	Whether stale memories are marked automatically (Find stale memories)	Off

• The line at the top shows whether memory is on, how many memories in how many projects, and when it last changed. When the sync with Claude Code needs you, “Syncing with Claude Code needs your attention.” and Review in Settings appear
• “Use Agent Fleet memory from outside Agent Fleet” is independent of the switch above. How to connect opens how to connect

What turning it on changes

• On takes effect from the next session started. Off refuses running sessions’ reads and writes at once
• Each agent’s instruction file gets a short fixed note that the memory tools exist and when to call them. It never holds memories. Turning the switch off removes that part only, without touching your own text
• Cursor has no place for an instruction file, so the same guidance is in the tool descriptions
• Local-model (lcpp) sessions start with the project’s memory index in their system prompt
• The Memories and Changes tabs work even with the switch off

Carrying Claude Code’s memory across

Claude Code keeps a memory of its own (MEMORY.md and the like). Flipping the switch carries its contents across by itself.

Flip	Confirmation dialog	What happens
Turn on	Turn on Agent Fleet memory? → Turn on and import	Imports, in the background, the memory of every Claude project that matches a working copy under ~/repos. Claude sessions started from then on run with Claude Code’s own auto-memory switched off
Turn off	Turn off Agent Fleet memory? → Turn off and write back	Takes a snapshot of Claude’s memory first, then writes back to Claude Code without overwriting anything. Claude Code’s own memory stays off until that has finished

• The dialog shows how many memories are imported or written back, and how many files are skipped because they look like secrets. While the workspace is stopped there are no counts; the work runs when it starts
• The automatic import never overwrites a memory you edited here with a Claude file: it only lists it as a conflict. A memory you made agents forget is not brought back either
• Claude sessions started before the switch keep using their own memory. When the last of them ends, what they saved is imported once more
• Under the switch, the page shows the progress, the last result and the files that need you (conflicts, secrets). One notification tells you when files were left
• The work carries on across a restart of the workspace

Codex and Muse Code

Agent	While on	When switched back off
Codex	Switches codex’s own features.memories off. Nothing is imported or written back	Puts back the earlier value (never turns it on by itself)
Muse Code	Starts Muse Code with its own memory switched off. This wins over the setting in your own environment. Turning the switch on also imports Muse’s existing notes, as below	Changes nothing; Muse keeps its default. Nothing is written back into Muse’s stores

For both, only sessions started after the flip are affected. A running session keeps what it started with.

What is imported from Muse Code

Turning the switch on runs the same import as Import from muse, in the same background run as Claude’s and with the same rules: secrets are skipped, a memory you made agents forget is not brought back, and a note whose Agent Fleet version differs is never overwritten.

Muse store	On turning the switch on
Project notes (.agents/memory in a working copy)	Imported into that working copy’s project. They sit inside the repository and may be committed to git
Personal project notes	Imported into the project of the matching working copy
Personal notes	Not imported. They would be visible to every project, so the page lists them as “needs your confirmation”. Import them yourself with Import from muse and the user-wide option

• The dialog also counts the Muse notes it leaves alone and the personal notes that are waiting. Under the switch, a Muse note whose Agent Fleet version differs is listed as “Agent Fleet has its own version, left as is”
• Switching on twice imports nothing twice. A Muse folder that cannot be read is counted and skipped; it never stops the switch
• Muse sessions already running keep using their own memory, and what they save is not imported again by itself. Use Import from muse for that

When it goes wrong

The write-back keeps failing

“The write-back keeps failing” appears under the switch, with the reason. Until it succeeds Claude Code’s own memory stays off, and Agent Fleet retries with growing pauses.

• Retry now tries again at once
• Switch Claude Code’s own memory back on anyway gives Claude Code its memory back without waiting for the write-back. What Agent Fleet holds does not appear on the Claude side until a write-back succeeds

Memory is on, but sessions do not use it

On takes effect from the next session started. Start a new session.

Related

• Read and correct agent memory: read and correct memories in the Memory view
• Use memory tidy-up: the Memory tidy-up (suggestions) switch
• Find stale memories: the Automatic lint marks switch