Restore memory from a snapshot

See the history of the agent’s own memory (what Claude Code, Codex and Muse Code keep in their own files), and put it back to how it was at a point in time. “It learned something it shouldn’t have” and “when did this go wrong” become fixable after the fact. You can also take it to another environment.

How it differs from Agent Fleet memory is in Choose which memory your agents use. Check and revert Agent Fleet memory’s changes in the memory Changes view.

How it works

What you do	What happens
Nothing, day to day	A few minutes after an agent stops, a snapshot of its memory is taken
Pick a point in the history and restore it	The memory is replaced by its contents at that point. The state just before is snapshotted too
Export, then import in another environment	The memory and its history move to the other environment

Open the screen

Open ⚙ Settings → Memory snapshots. It works only while the workspace is running.

At the top are the time of the last snapshot, whether automatic snapshots are on, and “Snapshot now”. Below are three views, History, Targets and Export & import. The view you leave open is remembered.

See the history

History lists the snapshots newest first. Select one to see the diff of what changed at that point.

Trigger	When it is taken
auto	A few minutes after an agent stops
manual	When you press “Snapshot now”
pre-restore	Just before a restore
restore	When you restore
pre-write-back	Just before Agent Fleet memory is written back to Claude Code
import	When you import an exported file
agent memory	A write to Agent Fleet memory (hidden by default)

• Find a time: enter a date and time in Point in time and press “Go to this time” to select the snapshot of that moment
• Agent Fleet memory writes are hidden, because the memory Changes view already lists them. Each run of them is folded into one line, “Agent Fleet memory writes hidden: N”, and “See them in Agent memory > Changes” opens them. To list them, tick “Show Agent Fleet memory writes” (remembered on this device)

Restore a point in time

1. In History, select the snapshot of the point you want back
2. Press “Restore this point”
3. Choose what to restore: “Everything”, or “Pick what to restore” and pick agent kinds (Claude Code, Codex, Muse Code) or projects
4. Press “Restore”, then “Restore” in the confirmation

• A restore can be undone too. The state just before it is snapshotted automatically (trigger pre-restore)
• When a session is running, the confirmation warns you. Anything it writes after the restore shows up as a new snapshot in the history

Check the targets

Targets lists the memory being versioned, per agent kind, with its file count, size and projects.

• Codex’s memory is off by default. Turn it on here to version it. While it is on, Codex keeps spending tokens to build memory from its conversations
• Turning on Agent Fleet memory switches Codex’s memory off once. If you turn it on here after that, it runs beside Agent Fleet memory and stays on when Agent Fleet memory is turned off
• Muse Code’s memory is versioned when it exists: the notes under ~/.local/share/muse/memory/ (its personal and personal-project stores), restored as one kind. Muse Code’s project memory (.agents/memory inside a repository) is not versioned here, because it lives in the repository and git is its history
• Restoring Muse Code’s memory while a Muse Code session is running is allowed: files are replaced one at a time and never half-written, and Muse Code’s lock file is left alone. Anything it writes after the restore shows up as a new snapshot
• The Automatic snapshots switch decides whether snapshots are taken by themselves. Nothing is stored when nothing changed. On some deployments the operator has disabled automatic snapshots

Export

In Export & import, export the memory to a file.

1. Choose the format

   Format	Contents
   bundle (full history)	Every snapshot. Agent Fleet memory is included
   tar.gz (latest only)	Only the current contents

2. Press “Export”

If anything looks like a secret, you are shown only its first few characters and asked to confirm. Once you have checked, press “Export anyway”. The exported file is not encrypted. Review the contents before sharing it.

Import

Import a file exported from another environment. Selecting a file alone changes nothing.

1. In Export & import, choose a bundle or tar.gz file

2. Check the format, the number of snapshots, the latest time and the count of items that look like secrets

3. For a bundle, choose how to apply it

   How to apply	Result
   Replace only the selected scope (keep this environment’s history)	The memory of the projects or kinds you select is replaced with the imported contents
   Migrate with history (replace this environment’s history)	The imported snapshots become this environment’s history, so you can roll back to any of them. The scope is always everything

4. Select what to replace and press “Import” (“Migrate” for a migration)

• The state before the import is snapshotted, so an import can be undone later too. A migration stashes the earlier history, which no longer appears in the list
• A file up to 64 MiB can be imported. An administrator can raise that to at most 128 MiB (limits)
• Agent Fleet memory in a bundle is not written here. When a bundle carries some, “This bundle also contains Agent Fleet memory” is shown; use “Open Import & export” to import it separately

When it goes wrong

“The workspace is stopped” is shown

The history lives inside the workspace. Press “Start the workspace” and open the page again.

Automatic snapshots will not turn on

When the note under the switch says automatic snapshots are disabled for this environment by the operator, the operator has turned them off. You can still take one by hand with “Snapshot now”.

Codex’s history does not appear

Targets shows Codex as “Memories are off (Codex’s own default)”. Turn it on, and Codex is versioned from the next time it runs.

The import says “Cannot import: this environment has no destination for”

An agent kind’s own memory in the file is not in use in this environment (for example, Codex’s memory is off). That kind cannot be selected. For Codex, turn it on in Targets and import again. Muse Code’s memory has no such switch: it can always be imported, and its folder is created if it is missing.

The import says “The secret scan of the imported contents failed”

Nobody has confirmed there are no secrets. Check the contents yourself before you decide whether to import.

Related

• Import and export memory: import Agent Fleet memory, and write it back to Claude Code
• Use memory from an external agent: read and write Agent Fleet memory from outside the workspace