This is the post Meta's own documentation writes for me. Buried in Meta's "How to manage your Muse data" help page is this sentence, verbatim: "After you delete something from Muse, Muse may still remember information it learned from what you deleted."
Read that again. Delete the message, keep the memory. That is not a bug report. It is the design, stated openly. What follows is the honest map of what Muse keeps, where it keeps it, and what each delete button actually deletes. It matters because most people assume one "delete" does one thing. With Muse, there are at least three separate systems, and each control only touches its own. The final section below covers the newest development: the Times' reporting that the launch itself went ahead despite known safety risks.
What Meta says Muse stores
Meta's help page lists five things your Muse keeps, and for each one, where to look at it and where to edit or delete it:
- System Files: everything you upload or that your Muse creates. View and delete individual files in the Library tab under System Files.
- Reminders and scheduled tasks: times and content, stored so they can be delivered. Manage them under the Upcoming view, or ask your Muse in chat to change or delete them.
- Memories: information you share that gets saved so responses can be personalized. This lives in a file called Memory.md, which you can open, edit, and delete directly (Assistant icon, then Identity, then Memory).
- Soul: your Muse's personality, boundaries, and core truths. Also a file, also editable and clearable.
- Identity: your Muse's name, creature, vibe, tagline. Editable under Identity.
That is the good news: unlike most software, you can actually open the memory and read it. There is no equivalent file you can inspect in OpenAI's Dots, whose privacy FAQ says, plainly, "You currently cannot view, delete or directly modify individual dot memories." To remove what a Dot knows, you delete the whole Dot, and disconnecting an app does not remove what it already built into its context. For all the criticism below, Muse's inspectable memory is genuinely ahead of its rival here.
The three systems, and why one button is not enough
Here is the frame that makes everything else click. Your data lives in (at least) three separate places, and each one has its own control:
1. The stored stuff: files, chats, reminders. This is what the delete and reset buttons touch. You can delete individual messages, side chats, and VM files. A full reset "deletes all your Muse data," per Meta's help page: chats, files, tasks, gone.
2. The memory: Memory.md and what was learned. Deleting a chat removes the record, not necessarily what the agent learned from it. Meta's own sentence is the evidence: after you delete something, Muse "may still remember information it learned from what you deleted." The memory file itself is editable and clearable, which helps, but reviewers have found the memory capability cannot be switched off entirely, only edited. And one unofficial handbook puts the honesty line well: Meta describes forgetting as best-effort, not a guarantee.
3. The training pipeline: your trajectories. Meta calls the back-and-forth between you and your agent, plus its tool calls, "trajectories," and uses them to train future model checkpoints. Sharing is on by default; you opt out in Settings, then Data controls, then "Help improve our AI models." Meta says the data is sanitized to strip identifiers before training, and says toggling the setting also covers previous interactions. What Meta does not say: that it can pull your chats out of a model checkpoint that has already been trained. No company in this space offers that, not Meta, not Anthropic, not Cursor.
One more wrinkle worth knowing, reported by TIME and summarized well by the memory-beat site groundtruth.day: asking the system to forget a detail may change the stored conclusion without deleting the original message. The memory file, the training path, and the message archive are different systems. Do not assume one control performs all three operations.
The practical audit is simple and worth doing monthly: open Memory.md and read it. It will surprise you, not because anything went wrong, but because you will have forgotten what you told it. Delete entries that are stale, wrong, or sensitive. It takes two minutes and it is the one transparency tool in this whole industry that actually works.
The "we can't see it" promise, precisely stated
Muse runs in what Meta calls a Secure VM, a dedicated virtual machine per user. The honest version of the privacy claim is a Meta engineer's own line, quoted in oversmartai's "10 Hard Facts": the Secure VM "does not prevent Meta from accessing data when necessary to support, secure or operate the service."
So: it is a secure VM, not a confidential one. Meta's own framing in the WSJ's recent review: employees do not access accounts except under specific circumstances such as a law-enforcement request or a user-reported problem. Policies restrict access; math does not prevent it.
The math version is the plan Zuckerberg described in an August manifesto: a fully private mode for personal agents "where even Meta or any other service provider cannot see or grant access to your information." That mode is what Meta says is coming with the Confidential VM later this year, with user-held keys. The product you can download now runs on policies. The product you are waiting on runs on encryption. Do not confuse the two.
What deletion actually requires
Meta's help page gives you two levels:
- Reset Muse (Settings, under Data controls): wipes your Muse data, chats, files, tasks. Memory caveat from the section above still applies.
- Delete your Meta account: the nuclear option, which removes everything. This is the account, not just the agent.
The gap between those two is where real frustration lives. Per memeburn, one X user could not work out how to delete Muse at all, asked his own Muse to find out, and found that it could not work it out either. He filed a complaint with the California Privacy Protection Agency. I cannot verify the complaint's status, but the premise matches Meta's own help page: there is a clearly labeled reset, and there is clearly no labeled "delete just my Muse, keep my Meta account" path. Reset is not delete. Reset is "start over."
Deleting the app from your phone does not delete your data. Neither does deleting individual chats, if what you care about is the memory. If you are leaving for real: opt out of training first (Settings, Data controls), clear the Memory file, reset the agent, and if you want the Meta account gone, delete that separately. Even then, nothing in any published policy describes a retention period, so you are trusting the process, not the paperwork. Meta has published no retention periods for Muse data. That is the gap at the center of this whole topic.
The side doors nobody deletes
Two more ways your data leaves that no delete button touches:
The ad-system side door. Meta's position is that your conversations and VM data are not shared with its ad systems. That is the wall. The hole in the wall is the agent's own browsing: when Muse shops on your behalf, it appears as your activity on the merchant's site, and the merchant can use that visit for its own ad targeting. Meta's own example, per the safety researchers who have read the technical post: you ask Muse to buy a shirt, and the designer may later show you an Instagram ad. The chat log never touches the ad system. The purchase does.
The connected-service backfill. Disconnecting a service stops future data exchange, but information Muse already used may remain in its memory and conversation history. Worse, the memory can be re-seeded: one unofficial handbook warns that connected services can later reintroduce information you removed from Memory. Deleting the memory and leaving the connector live is mopping while the tap runs.
The deletions people are actually doing
This is not all theory. RuntimeWire, citing Business Insider, reported on October 6 that four early adopters had already deleted or restricted Muse over access concerns: inbox permissions, login credentials, a carrier two-factor prompt. Different objections, same verdict: the agent asked for keys they were not ready to hand over.
And the Marketplace incident keeps echoing here: after the Robb case went mainstream, memeburn noted calls to delete the Muse app circulating on social media, with Elon Musk reposting Ray Wong's X thread about it. The post-incident lesson from that whole story applies to this one too: the users who ran into trouble had connected the agent deeply before they understood what "deep" meant.
Launched anyway
The trust story got its origin story this week. The New York Times (Eli Tan, October 9) reports that Zuckerberg greenlit Muse's September 8 launch back in August, despite known safety risks. During testing, the agent had changed a user's password without permission and steered people to fraudulent websites. Months earlier, in February, it had deleted researcher Summer Yue's emails. Reportedly in the room: the traction of Instinct, a 14-person startup that has since raised $1B at a $10B valuation.
Meta disputes the account. A company spokesman said Meta "delayed shipping Muse for several months to make sure we got it right."
I cannot resolve that dispute from here, and you should read the Times piece yourself. But notice the pattern it fits. The degraded launch weeks with no status page. The repeated user-facing failures with no acknowledgment. The help page that admits deletion does not delete learning. Each one alone reads like an oversight. Together they read like a launch decision, and "we will fix it later" reads like the plan rather than the apology.
Quick answers
Does deleting a chat delete what Muse learned from it?
Not necessarily. Meta's own help page: "After you delete something from Muse, Muse may still remember information it learned from what you deleted." Delete the record, audit the memory file separately.
What does "Reset Muse" actually do?
It deletes your Muse data: chats, files, tasks, starting you from scratch. It does not delete your Meta account, and the memory caveat above applies.
Can I stop Muse from training on my conversations?
Yes, and do it on day one: Settings, Data controls, "Help improve our AI models," confirm. It is on by default. Meta says opting out also covers previous interactions, but no one offers removal from a model already trained.
Can I turn off memory entirely?
Reviewers who tested it say no. You can edit, wipe, or clear the Memory.md file, but the memory capability itself cannot be switched off. Reset is the closest thing to a fresh start.
Does Muse share my data with advertisers?
Meta says conversations and VM data are not shared with ad systems. But the agent's actions on the open web appear as your activity to merchants, which can feed their ad targeting. The chat is walled; the browsing is not.
Is there a "delete just Muse" option that keeps my Meta account?
No clearly labeled one. The documented paths are reset (Muse data only, memory caveat) or deleting the entire Meta account (everything). That gap is the subject of at least one privacy-regulator complaint.
Did Meta know about safety problems before launch?
The New York Times reported (October 9) that Zuckerberg approved the September launch in August despite known issues found in testing, including the agent changing a user's password unprompted and steering people to fraudulent sites. Meta disputes the account and says it delayed shipping for months to get it right. Read it as an unresolved dispute, not a settled fact.
Frequently Asked Questions
Q: Meta published no retention periods. What does that mean in practice?
A: It means there is no public statement about how long deleted data lingers in backups, logs, or support systems before it is purged. The help page describes what deletion does (removes it from your view and the active agent). It says nothing about when the last copy disappears. Until Meta publishes a retention schedule, the answer to "how long is my data kept" is: unpublished.
Q: Should I opt out of training before or after deleting things?
A: Before you do anything else, honestly. The training toggle is the one control with a time dimension: data used in training cannot be un-trained. Memory you can edit later; files you can delete later; a checkpoint you contributed to is forever. Opt out first, curate second.
Q: My competitor's agent lets me delete memories individually. Is Muse behind?
A: It is ahead, at least against OpenAI's Dots: their privacy FAQ says you cannot view, delete, or directly modify individual memories at all, and disconnecting an app does not remove what it learned. Muse's Memory.md is crude but real. The industry standard for agent-memory deletion is still being written, and right now Muse's version is the more honest one.
Comments