Read and correct agent memory

Agent memory is Agent Fleet’s own memory, shared by every kind of agent. This page shows how to read what agents remembered, rewrite it, pin it and make agents forget it. How it differs from instruction files and the agent’s own memory is in Choose which memory your agents use.

How it works

What you do	What the agent does
Let sessions use memory in Set how memory is used (off by default)	Reads the index when it starts work, and searches and reads the memories it needs
Read, correct and pin memories in the Memory view	Saves what it learned with `memory_save`, and fixes part of one with `memory_patch`
Make agents forget a memory you do not want	A forgotten memory disappears from every agent’s view
Check the changes and revert bad ones	Reports a memory that was stale or wrong with `memory_feedback`

Memory has one scope per project plus a “user-wide” scope that every project receives. A worktree shares its repository’s memory
What one agent learned survives a switch of agent kind, a handoff, and a child session of another kind
An agent’s save is shared at once, without your approval. A save that looks like it holds a secret is refused

Open the Memory view

Open ⚙ Settings → Agent memory and press Open memory view at the top. The Memory view opens in a pane of its own (the Memory tab), with three tabs: Memories, Changes and Suggestions.

You can also search in the command palette’s Memory scope and use Open in Agent memory (Use the keyboard and the command palette)
There is one Memory view per workspace. Opening it again moves you to the one already open
Memory lives inside the workspace, so you can read it only while the workspace is running

The Memories tab of the Memory view. On the left the scope, the filter chips, the search box and the list of memories; on the right the selected memory's text with Edit, Pin and Forget this memory

Find and read memories

Above the list on the left, pick a scope and a chip.

What you pick	What the list shows
Scope All scopes (the default)	The user-wide memories and those of every project. Each row says where it lives
Scope User-wide (shared with every project), or one project	Only that scope’s memories
Chip All, Pinned or Forgotten	Everything, the pinned ones, or the forgotten ones

The Search memories box looks at once for every word you typed (in any order) in the name, description, type and author. When you stop typing it searches the bodies too
When no memory holds every word, it shows the ones that hold some, under “No memory contains every word. Showing memories that contain some of the words only.”
Pick a row and the right side shows the text, scope, type, Author, Updated and Uses (how often agents read it). Your own reading and searching do not count as uses
A memory whose name or text looks like a secret is not listed, only counted
On a phone the list and the detail are two screens. Back to the list (or Escape) returns from the detail

Write and correct memories

New memory creates a memory in the scope on screen. While all scopes are shown it goes to the user-wide scope
Edit in the detail lets you change the name, description, type and text, then Save
A save goes through the same revision check and secret scan as an agent’s save, and is recorded in Changes as written by You (Console). You can revert it later
You can write into a project only once that project already holds memory
An edit does not change which kinds of agent may read the memory (when it is limited to some)

Pin a memory

Press Pin in the detail and the memory is always listed first in the index agents read when they start work. Unpin takes it back.

Memories that are not pinned are ordered by type (feedback and user first), then by how often they are read, then by recency
The index has a size limit (about 24 KiB of described lines), and pins share it. If the pins alone do not fit, the last of them drop to the names-only list. Pin only what every session needs
Only you can pin and unpin. An agent’s later update of a pinned memory keeps the pin
A pin or unpin is recorded in Changes as “pin changed” and can be reverted

Forget a memory

Press Forget this memory in the detail and no agent sees the memory any more. Its text stays in the history.

Forgotten memories are listed under the Forgotten chip, newest first. Unforget, on the row or in the detail, brings back the text it had just before
If that earlier text now trips the secret scan, Unforget is refused. Use Revert this change on its forget entry in Changes, where you can check the text before bringing it back
A memory with no change in its history, such as a file put in place by hand, cannot be forgotten from here
History of this memory, below the detail, lists that memory’s own changes. Pick one to see its diff. If older changes are missing, press Load older changes

Remove past text from the history

Forgetting only hides a memory: its past text stays in the history and in exports. To delete past revisions from the workspace, for example after writing a secret into one:

In the memory’s detail, open More actions (⋯) → Purge history…. For a forgotten memory it is also on the row and in the detail
The dialog shows how many change records go, which histories are rewritten, and what the secret scan found in past revisions (values masked)
Type the memory’s name to confirm, and press Purge history

The current text stays. To remove the current text too, edit the memory and save it first, then purge
A forgotten memory loses every revision, the forget record included, and can no longer be unforgotten
The memory’s earlier changes can no longer be reverted. Other memories are not affected
Exports you already made cannot be recalled. Delete downloaded bundles or archives, and copies made elsewhere, yourself
A purge can only be started from this screen. A session cannot start one

What agents see of a memory

A memory has a status that changes how agents see it. Lint, agents’ reports, the tidy-up and you set it; an agent cannot set it itself.

Status	How agents see it
Normal	In the index, in search and when read
Suspect (has reasons)	Wherever it appears, its reasons come with it (Find stale memories)
Quiet, or of kind `log`	Not in the described part of the index. In the names-only list, in search and when read
Archived or superseded	Not in the index and not in search. Reading it returns a stub with the reason instead of the text

Your own Memory view shows every memory, whatever its status
Reverting a change brings back the earlier version whole, its status included

When it goes wrong

Saving says “The text looks like it contains a secret, so it was not saved”

The text or the name tripped the secret scan. Remove that part and save again.

“Try again in a minute” appears

For about half a minute after the workspace starts, saving, pinning and forgetting are not accepted. Wait a minute and press it again.

“cannot be forgotten from here” appears

A memory with no change in its history (a file put in place by hand, for example) cannot be forgotten from here.

Set how memory is used: the switch that lets sessions use memory
Check memory changes: who changed what and when, and how to revert it
Find stale memories: check for broken links and old paths
Use memory tidy-up: have a model draft fixes