Memory

Mention once that you like tabs, or that tests live in spec/, and the agent can carry it into every session after. It asks before it saves a note, and this screen shows you every one it keeps: “What the agent remembers about you, across every workspace.”

Where to find it

Abilities → Memory in the sidebar.

What’s on the screen

Global memories

“These apply in every project.” Global means in every workspace. These are notes about you: how you like to work, things you’ve corrected. Notes about a single workspace live in its workspace settings, and, when a workspace is open, the line under the list links there.

The agent keeps each set of notes as ordinary text files on this computer. Nothing is sent anywhere or shared with anyone you work with.

The cost line

The first line reads, for example, “Costs ≈ 120 tokens every turn. Memories open on demand. 12 of 100 used.”

The cost is in tokens (the unit models read and bill text in, roughly four characters; more on context). On every turn the agent is shown a one-line summary of each note, and that’s all memory costs until it opens a note, which it does only when the summary looks relevant. The number is the same estimate the session’s context panel shows.

Each set holds up to 100 notes. When one is full, the agent has to merge or drop a note before it can add one.

The list

Each row shows what kind of note it is, its name, how long ago it changed, and its summary. The kinds are:

• About you
• Something you corrected
• About this project
• A pointer to something

Once you have more than four notes, a Search memories box appears above the list.

With no notes yet, the screen says No memories yet, then: “Tell the agent “remember that …”, or it will offer to save what it learns from your corrections. Every save asks you first.”

Forget all… and Import from Claude Code

Import from Claude Code sits under the list. Once you have a note, Forget all… sits there too. Both are covered below.

Notes can be left behind when you remove a workspace from the sidebar, or when its folder moves into or out of git. A banner then says how many, with a button for each leftover folder, like Forget 3 from 1a2b3c…. Click one to delete that folder’s notes.

How memory works

The disclosure at the bottom explains the rules in full. Two worth knowing. Notes are hints, not facts: the agent is told to check anything that might have changed before it acts. And it won’t save anything that looks like an API key, an access token or a private key; that’s refused with a reason.

How a memory gets saved

You say “remember that …”, or the agent offers to save something it learned from a correction. Either way, it asks you first.

1. The agent asks to save. An approval dialog opens: “The agent wants to remember this”.
2. Read the note itself: its Name, Kind, Scope and Summary, then what it remembers. The scope says where it applies: “About you, in every project” or “About this project only”.
3. If it would replace a note you already have, the title reads “The agent wants to REPLACE this memory”, and you see what changes instead of the whole text.
4. Click Allow to save just this one, or Deny to keep it out. Always allow means future saves stop asking, so pick it only if you’re happy for the agent to save notes on its own.

Forgetting asks the same way: “The agent wants to forget this”. Opening a note to read it never asks, since you can read every one on this screen anyway. More on the approval dialog and its other buttons in Approve a tool call.

Read, edit or forget a memory

1. Click a note in the list. It opens in full, with where it came from, for example “Saved by the agent in “Fix the login bug””, or “Edited by you, or imported”.
2. To change it, click Edit. You can change the Summary (“this is the line the agent sees every turn”, up to 150 characters) and What it remembers. Click Save, or Cancel to leave it as it was.
3. To delete it, click Forget. The app checks: “Forget “name”?” Click Forget it, or Keep it to change your mind.

Forget everything

1. Click Forget all… under the list.
2. The app asks, for example, “Forget all 12 global memories? This cannot be undone.”
3. Click Forget them all, or Keep them.

This only clears the set you’re looking at. A workspace’s own notes stay until you clear them in its workspace settings.

Import from Claude Code

If you’ve used Claude Code on this computer, you can bring its memories across.

1. Click Import from Claude Code. The dialog explains: “Memories are copied in, never linked. Anything already here with the same name is skipped and named.”
2. The memories are grouped by the project folder they came from. Untick any you don’t want.
3. For each folder, pick where they go: Into global memory, or Into this project when a workspace is open. “The destination applies to a whole folder.”
4. Click Import selected. The dialog tells you how many came in, and names any it skipped, with why.
5. Click Done.

If there’s nothing to import, it says “No Claude Code memories found on this computer.”

Turn memory off

Memory has no switch on this screen. It lives on Built-in tools, as Memory — 3 tools. There you can also read the memory policy the agent follows and add your own lines to it. When it’s off, this screen says “Memory is turned off for this app.” Your notes stay where they are.

To turn it off for one workspace only, untick Use memory in this project in its workspace settings. Global memories still apply there.

During a session

• A save or a forget waits for you in an approval dialog, showing the note itself.
• The session’s context panel counts memory’s cost, with the same number as this screen.

Related

• Built-in tools: memory’s on/off switch and its policy.
• Approve a tool call: how the approval dialog works.
• System prompt: standing instructions for every session, when you’d rather write them yourself.