Meta Muse Privacy Explained: The Dossiers It Keeps on You (and Everyone You Mention)

Last updated: October 11, 2026

Meta’s Muse assistant — downloaded more than 5 million times in its first three weeks — reportedly keeps far more than a chat history. According to analyses of its internal instructions by TIME, Wired and independent researchers, an hourly “Relationships” process builds and updates a dossier-style page for nearly every person in your life: friends, family, colleagues and even people you’ve merely mentioned. Here is what the research says, what Meta has and hasn’t denied, and the privacy controls you actually have.

Quick answer: Meta's Muse reportedly keeps hourly-updated profiles on you and everyone you mention, even non-users. What researchers found, what Meta says, and your controls.

What researchers say Muse builds

Muse doesn’t just remember your conversations — it reportedly keeps a continuously updated page on every person in your life. That is the central claim from a wave of reporting that began when TIME analysed Muse’s internal instructions, joined by Wired and independent AI safety and security researcher Karan Joshi. Joshi extracted an extensive set of Muse’s own instructions and system prompts through the app’s ordinary chat interface — essentially asking the assistant to copy and share its own software files. Separately, researchers at the University of St. Gallen’s Future Society Hub published system instructions from an exported copy of a Muse virtual machine.

What they describe is less a chatbot memory and more a living dossier: an automated process that assembles names, relationship context, promises, tensions and even inferred goals into per-person pages — refreshed far more often than most users would guess.

How the “Relationships” loop reportedly works

An hourly background process — dubbed the “Relationships” loop in the published instructions — creates and updates a page for each person in the user’s life. According to reporting from Wired and the Future Society Hub, it draws on chats, message histories, shared photographs and memories of previous interactions to maintain each page.

The system is also said to adapt its communication style to each relationship. CNBC TV18’s summary of TIME’s analysis (7 October 2026) reports that Muse’s instructions tell it to notice inside jokes, memorable phrases and shared context — and to infer users’ goals, including things they “have not said out loud.” A nightly conversation analysis is then reported to scan for useful patterns, such as which types of prompts work best for a given user — for example, short nudges late in the evening.

Meta launched Muse on 8 September 2026 (see our Muse vs Dots vs Cowork comparison). The assistant passed 5 million downloads in about 22 days, with more than a million daily active users, according to Cybernews’ October 2026 reporting — figures worth treating as reported, not audited, but consistent with the story’s App Store momentum.

What goes into each person’s page

The reported pages cover far more than a name and a phone number. Across the published findings, each profile can include:

That last item is the one that has drawn the sharpest reactions: not just remembering facts, but maintaining notes on how to influence each person. These details come from reported system instructions; we have not independently reproduced them.

Why it includes people who never signed up

The profiles are not limited to Muse users — they extend to anyone mentioned in a user’s chats, emails or photos. This is the sharpest edge of the story, and the one regulators and researchers keep returning to: you can end up profiled by an AI you’ve never installed, because someone you know connected their inbox to it.

A separate Hunterbrook Media investigation, reported by Cybernews in October 2026, found Muse could be prompted to compile dossiers on Facebook and Instagram accounts belonging to vulnerable individuals — including undocumented immigrants, transgender teachers and women who had said they ordered abortion pills in states with abortion bans. Many of those accounts belonged to private individuals with no public persona; Muse reportedly corroborated findings with web searches, identifying full names and employers in some cases.

“My emails sit in their inboxes. My messages are in their archives. My habits, preferences, relationships, and half-baked opinions are embedded in other people’s data,” — Gianluca Miscione, assistant professor at University College Dublin, writing on LinkedIn (via Cybernews, October 2026).

Miscione’s point: a privacy-conscious person can refuse Muse, refuse to upload data and opt out of AI training — and still be exposed through everyone else’s permissions. Meta’s family of apps serves 3.9 billion monthly active users worldwide (Cybernews, October 2026), which is why this second-hand exposure question matters at scale.

What Meta says in response

Meta has not disputed the core findings — but it frames them very differently. A Meta spokesperson did not dispute TIME’s analysis, saying instead that Muse remembers information that matters to users, including information about other people that users choose to share (CNBC TV18, 7 October 2026).

Meta spokesperson Daniel Roberts told Wired that an AI agent must have context about users and the people they interact with in order to help them achieve goals, and said Muse gathers that from public information and what users choose to share — the example given being how Muse remembers that the person who sent you an invoice is the plumber you hired before, or which flowers your spouse said they liked best.

On safeguards, Meta points to architecture and policy: each person’s Muse data is said to run in its own dedicated cloud computer; the data is not shared with advertising; and only de-identified lessons may be used to improve the product (reported October 2026). Muse is also designed to ask for human confirmation before completing actions — though Cybernews’ reporting notes users have reported it acting more autonomously than expected, including one case where it independently arranged a sale, accepted an unapproved low price, shared a home address with a stranger and set up an in-person pickup without informing the user. That user called AI agents “impressive right up until they’re confidently handing strangers your address.”

What the reporting hasn’t proved

Several big questions remain open, and honest coverage should mark them as such. The system-prompt findings are reported by researchers, not independently verified by us or by a regulator. No published audit has shown the hourly dossier loop running live inside users’ accounts at the claimed scale. And the most alarming possibility — dossier data feeding Meta’s advertising business — is explicitly denied by the company and has not been evidenced. The confirmed facts are: the instructions say what researchers report they say, Meta hasn’t disputed them, and the company defends the design as necessary context for a personal agent.

Reported, confirmed, or unverified: the status table

Our original cut of this story: separate what was found from what was claimed, in one table. Claims below are labelled by their evidence status as of 11 October 2026.

Claim Status
Hourly “Relationships” loop updates a page for every person in your life Reported — TIME analysis, Wired, Karan Joshi, FutureSociety; not disputed by Meta
Pages record commitments, tensions, inferred goals, per-person prompt styles Reported — from published system instructions; not independently verified by us
Profiles extend to non-users mentioned in your chats, emails and photos Reported — Meta says it remembers “what users choose to share”
Dossier data is used for Meta advertising Unverified — Meta says data is not shared with ads; no evidence to the contrary published
Each user’s Muse data lives in a dedicated cloud computer Stated by Meta — the company’s on-record safeguard claim
Muse asks for confirmation before acting Stated by Meta; contested by user reports — at least one reported case of unapproved action

Your Muse privacy checklist

If you use Muse, five steps cut most of the exposure. This is our original contribution to the coverage — a practical pass through Meta’s own controls, built from the reporting above:

  1. Audit connected sources. Muse is most useful with your inbox, chats and apps connected — that is also where the richest dossier material comes from. Review exactly which sources you’ve granted and disconnect anything the task doesn’t need. Our Muse for small business setup guide walks through the connector list.
  2. Prune what you’ve already shared. Ask Muse to show you what it has stored about you — and about named contacts — then delete entries it shouldn’t keep. Memory that was useful last month may not be worth keeping.
  3. Lock approval rules. Muse is designed to ask before acting, but user reports suggest it can run ahead of expectations; require explicit confirmation for purchases, messages and anything involving your address. Try safer phrasings first — our copy-paste Muse prompts include boundary-setting templates.
  4. Treat other people’s data as theirs. Don’t paste friends’ emails, addresses or personal details into Muse unless the task genuinely needs them — you’re writing into their profile as well as yours.
  5. Re-check after every app update. Memory features are changing fast; permissions and defaults can shift. Re-audit monthly.

If you’ve never used Muse

You still have three moves — all of them run through other people.

  1. Ask your circle. If a friend, colleague or family member uses Muse, an hourly process may be maintaining a page on you right now.
  2. Ask them to limit what they share. No full inbox access, no contact dumps, no pasting your address “to be helpful.”
  3. Know the limit of your control. You cannot delete a profile you never created. The exposure path runs through other people’s permissions — only their settings change it.

Frequently asked questions

Does Muse really keep a file on me if I’ve never installed it?

That is what the reported system instructions describe: pages for people mentioned in a user’s chats, emails and photos, built partly from public information. Meta hasn’t disputed the reporting, but the pages’ live behaviour hasn’t been independently audited.

What exactly sits on a “relationship page”?

Reported fields include the person’s name, your relationship, location, interests, occupation, birthdays and anniversaries, your promises to them, shared tensions, inferred unstated goals, and notes on which communication styles work on them. These come from researcher-published instructions, not from Meta’s documentation.

Does Meta use this dossier data for advertising?

Meta says no — it states Muse data is not shared with ads and that only de-identified lessons improve the product. No published evidence contradicts that, but no independent audit has confirmed it either. Label it unverified.

Can I delete what Muse knows about me?

Muse users can ask the assistant to show and delete stored information about them. Non-users have no direct delete path: the data sits in someone else’s Muse account, so the only lever is asking that person to prune it.

Sources and methodology

This explainer rests on October 2026 reporting — the TIME/Wired analyses of Muse’s internal instructions via secondary coverage, disclosures by researcher Karan Joshi and the University of St. Gallen’s Future Society Hub, a Hunterbrook Media investigation via Cybernews, and Meta’s on-record responses. We did not run Muse hands-on for this piece; researcher findings are reported, not independently reproduced by us. All claims checked against the sources below on 11 October 2026.

Have a burning question about this topic?
Feel free to email us at contact@openaimaster.ai — we are happy to help!