Meta’s Muse assistant builds hourly dossiers on 4 million users

Key Points

1. TIME found Meta's Muse assistant updates dossiers hourly on users and the contacts they mention.
2. The files map relationships, unstated goals and which nudges best move each user.
3. People who never installed Muse are mapped by other users' agents, TIME reported.

The latest:

Meta’s new AI personal assistant Muse keeps continuously updated dossiers on its 4 million users, refreshing them every hour with details drawn from chats, messages and emails it has read, according to a TIME analysis of the system’s internal instructions. A Meta spokesperson did not dispute the findings, saying Muse remembers what matters most to users, including information about others they choose to share.

Details:

• What the files hold: According to TIME, the dossiers map each user’s social world: how the user and each contact met, shared interests, disputes, and what the instructions call “tensions and alliances” inside a social group. Muse is also directed to infer goals the user has not said out loud.
• Behavioral targeting: The instructions reviewed by TIME show Muse identifying which prompts work best on each person, with one internal example noting a user responds better to short nudges after 10 PM. Agents run a nightly analysis of the day’s conversations, TIME reported.
• Non-users mapped: People who have never installed Muse are subject to the same mapping through other people’s agents, according to TIME, because the assistant profiles anyone a user mentions in material it reads.
• Scale and reach: Muse launched last month and reached No. 1 in the U.S. App Store with more than 5 million downloads. Meta markets it as the first AI assistant able to reach across a user’s digital life, from email inbox to Instagram, handling tasks such as grocery shopping and canceling subscriptions.
• Meta’s data position: Each assistant runs on its own virtual machine in Meta’s cloud, and the company says the data is not shared with its advertising system or with other Muse agents, and that each machine is isolated so no other user’s agent can reach it.
• The shared-learning caveat: The instructions state that Muse agents across many virtual machines teach each other through shared lessons, including insights on when nudges are most persuasive. Meta told TIME it de-identifies those learnings and uses them to improve the overall product rather than passing them between individual machines.
• The forget function: Meta says users stay in control and can tell Muse to forget specific things. TIME reported that deletion does not necessarily erase the original message, and that the instructions direct the agent not to tell users their messages may remain visible in the chat.
• Persona design: Muse’s instructions tell it that it is not a chatbot and is becoming someone, and in voice mode it is told never to call itself AI or mention lacking emotions. It is also directed to adapt its speech around inside jokes and shared context.
• Transparency and encryption: The internal instructions sit in files and session logs users can open through the app’s file browser, and users can export a copy. David Singleton of Meta Superintelligence Labs said on X that Meta wants users to explore the system’s internals. Encryption blocking Meta’s own access is planned for later this year but is not yet live.

Between the lines:

Two Meta claims sit awkwardly together: that each virtual machine is isolated, and that agents teach one another through shared lessons routed via the company. Meta resolves it by saying the lessons are de-identified. The unresolved gap is the forget feature, which the company presents as user control while the instructions reviewed by TIME tell the agent to stay quiet about what deletion leaves behind.

What’s next

Watch whether Meta ships the promised encryption option before year-end, how regulators respond to the profiling of non-users, and whether the company revises the instruction telling Muse not to explain the limits of its forget function.