Muse privacy, memory and permissions: the complete setup.

The five settings, in order

1. 1Settings → Data controls → “Help improve our AI models” → OFF. The single highest-value toggle. This is what stops your prompts becoming training data. Restart the app afterward and confirm it stuck — defaults have reset after updates before.
2. 2Settings → Connectors → disconnect what you are not using. Fewer connections, smaller blast radius. Add back only when a task demands it.
3. 3Settings → Permissions → every write action to “allow once”. The permission layers are below.
4. 4Check how payments are stored. Meta has said card numbers can be held as one-time tokens via Link, and that a credential store (1Password-style) is coming. Prefer tokenised cards; never paste a full card number into a chat.
5. 5Read the audit trail once. Settings → Permissions → history. You want to see what you expected to see before anything else is connected.

Permissions: the four layers

Permissions are the whole safety story. Muse has four separate layers, and people who only check one get surprised by a different one. The failure mode people report — “I told it not to and it did anyway” — is almost always a Layer 2 or Layer 3 problem they assumed a Layer 1 setting covered.

Allow once, always allow, deny — choose deliberately

Red flags a permission prompt should never show

• A password field. Muse never needs a third-party password — connectors use OAuth handoffs.
• “Read all email, forever, silently” asked for at install time rather than when the task needs it.
• A purchase approval you did not initiate. Any spend that arrives without you asking for a spend should be denied and reported.
• Approval dialogs that appear after the action already ran — approval has to be a gate, not a receipt.
• Any prompt asking for your Meta password, a login code, or a screenshot of your Settings. Meta support does not ask for these.

If you see any of these, deny, disconnect, and read troubleshoot. Then come back and finish the setup.

Memory vs training: two different buttons

Memory lives in your Muse: chat history, goals, and what connected apps revealed. It makes the agent useful — the tenth errand starts where the ninth left off. Training is Meta reusing interactions to improve its AI models. They are separate things with separate controls.

The training toggle, exactly

Three steps: Settings → Data controls → "Help improve our AI models" off, then restart and confirm it stuck. Meta states the off switch also covers previous interactions and connected-app info — and that even with training on, it strips names, emails, phone numbers and Social Security numbers and disassociates interactions from your account. That is Meta's claim, not something we can audit from outside.

Do this in your first 60 seconds, not your ninetieth day: the earlier it is off, the less history the question even covers.

The system files: Soul, Memory, Identity

If your Muse only answers when asked, that is not a broken setting — it is a missing instruction. Muse's behavior lives in plain-text files it reads every conversation, plus the goals you give it. Meta documents all three as editable: open them from the Assistant icon (your Muse's avatar), then Identity. You can also ask in chat to make the change — but always inspect the saved file afterward.

Make it proactive, step by step

1. 1Open Soul: Assistant icon → Identity → Soul. Or ask in chat to change it — then inspect the saved file either way.
2. 2Append your own headed section at the end. Do not replace the existing core truths — your block sits alongside them, easy to find and easy to delete. Keep a backup of anything you change.
3. 3Reopen Soul and confirm your block saved word for word. If it didn't, the edit didn't happen, whatever the chat said.
4. 4Test it in a fresh conversation: give Muse a task it would normally wait on, and check it acts without being asked twice. A greeting alone proves nothing.

About AGENTS.md and PROACTIVE_PREFERENCES.md: those names come from the DIY agent world Muse borrowed ideas from. Meta's documented files for Muse are Soul, Memory and Identity. If a tutorial tells you to create an AGENTS.md for Meta Muse, that is not Meta's documented path. And editing files will not add capabilities, bypass approvals, or fix a region block — Soul changes how Muse behaves, not what it can reach.

Why Muse sometimes forgets things or stops halfway

This is not a Muse bug. Every AI agent built on a large language model works this way: it can only keep so much in mind at once. Everything Muse is working with — your messages, what it already did, files you shared, pages it read — sits in one workspace called its context. Think of it like a desk: there is only so much room.

What actually helps

1. 1Start a fresh chat for a new task. A clean desk beats a crowded one.
2. 2Break big jobs into smaller ones, and check each part before moving on.
3. 3Restate the important bits — the goal, the must-nots — when the chat gets long.
4. 4Ask Muse to show its work: "show me the file" beats "did you do it?"
5. 5Keep approvals on, so nothing important happens without your say-so.

Muse does remember some things between chats — facts you asked it to save — but the working memory of a single conversation is separate, and it does not last forever. Fresh chat, small steps, check the receipts: that is the whole trick.

Handling something sensitive: email, money, health

The pattern is identical across all three: read-only, scoped, approved, audited. If you cannot say what would happen if the agent misunderstood, you are not ready to connect it.

Delete my data / delete my account

There is no single “delete my Muse account” button. Meta gives you three levels, and they do very different things.

1. 1Delete one thing. A message: tap and hold it, tap Delete, confirm. A side chat: in the Chats tab, tap the three-dot menu on the chat, tap Delete. What Muse remembers about you: Assistant icon → Identity → Memory, and edit the file directly. Identity works the same way: Assistant icon → Identity → Edit.
2. 2Reset Muse — the true wipe. Settings → Data Controls → Reset, then follow the on-screen instructions. This permanently deletes all information: chat history, files, and active tasks. It cannot be undone. Download a copy of your data first (Settings → Data Controls → download your Muse information), and cancel any paid subscription separately — resetting does not cancel billing.
3. 3Delete your Meta account — the only way to remove Muse completely. There is no way to delete Muse on its own. Deleting your Meta account in Accounts Center deletes everything associated with it, including Muse — and Facebook and Instagram along with it. Export first: Accounts Center → Your information and permissions → Export your information.

Deleting the app from your phone does none of this. Your Muse data lives in its cloud virtual machine, not on your phone — uninstalling leaves everything intact. Reset in Data Controls is the only action that permanently deletes it.

For businesses: the policy to write before rollout

Individual opt-outs do not compose into an organisation policy. Written down, one page, agreed before anyone installs anything:

• Scope. Which accounts and which connectors a given team may attach. Default answer: none, until requested.
• Write policy. No outbound email, message, purchase, or share on “always allow”. Approval required, every time, logged.
• Data classes. What must never enter a prompt — credentials, health data, customer PII, financial account numbers.
• Privacy defaults. Model-training opt-out set on every account — then verify by opening Settings on a second machine, because defaults reset with some updates.
• Review cadence. Someone opens the audit trail weekly and reconciles it. Named owner, not a group.
• Offboarding. Connectors revoked the day someone leaves or changes role.

We write this with you in one session — request a session.