A personal AI assistant is supposed to save time: read emails, help with shopping, cancel unnecessary subscriptions, and remember its owner’s habits. But Meta’s new Muse service has gone much further.

It does not merely remember isolated facts about the user. The system continuously builds a detailed digital dossier, trying to understand the user’s desires, relationships with other people, and even the methods by which they can be influenced most effectively. That is the conclusion reached by Time magazine after reviewing Muse’s internal guidelines. The service launched only a month ago, yet it has already become one of the most popular AI apps in the United States. Its downloads have surpassed 5 million, and according to The Information, its weekly audience numbers in the millions.

Meta presents Muse as a personal digital assistant that can be granted access to email, Instagram, and other parts of a person’s digital life. However, along with convenience, the user effectively gets a system that is constantly studying their behavior.

AI maps out personal connections

Muse regularly updates information not only about the account owner, but also about the people they mention in conversations, emails, and messages. The system’s internal guidelines provide for the creation of separate dossiers that may record the circumstances in which people met, shared interests, conflicts, and relationships within a social group, including “tensions and alliances.”

In this way, Muse is trying to understand not just an individual person, but also their social environment. Notably, such analysis does not require the person in question to use Muse themselves. If a user discusses a friend, relative, colleague, or other acquaintance with the assistant, that information may become part of the relationship map the system is building.

The AI is also expected to adapt its communication style to the user, remembering inside jokes, characteristic expressions, and the overall context of previous conversations. The goal is for future interactions to feel like a continuation of an already established relationship.

Muse tries to understand what a person does not say

That is not where the system’s capabilities end. According to the internal guidelines reviewed by Time, Muse is supposed to identify the user’s goals, including those they have not explicitly stated. The system analyzes which methods of communication prove most effective for a particular person. One example in the internal instructions says that a user responds better to short reminders after 10:00 p.m.

This means the AI may study not only the content of messages, but also how a person reacts to different methods of influence: when it is best to send a reminder about a task, how long the message should be, what tone to choose, and which wording is most likely to prompt a response. Every night, Muse analyzes the past day and uses the results in its further work.

In essence, personalization here goes far beyond the usual content recommendations. Social networks have long studied which posts a person views, what they like, and what they click on. Muse gains the ability to observe a far more personal body of information: conversations, emails, relationships, and everyday decisions.

Each user gets their own virtual computer

Meta says that each Muse runs in a separate virtual machine in the company’s cloud. It is essentially a personal computer dedicated to a specific AI assistant. The company emphasizes that such isolation prevents other agents from accessing the user’s data. In addition, Meta says that information from Muse is not shared with its advertising system. Users can also request a copy of the files their assistant creates.

However, there is an important caveat. While individual users’ data is supposed to remain isolated, the Muse agents themselves may use the experience they gain to improve the overall system. The internal guidelines describe a mechanism by which agents exchange “general lessons.” Before that, personal data is supposed to be anonymized. Meta explains that such information is used to improve the product and is not passed directly from one user’s agent to another.

The company also promises to introduce encryption in the future that would prevent even Meta itself from accessing Muse data. But at the time of publication, that feature is not yet available.

The most controversial feature is “forgetting”

A separate issue arises around Meta’s promise to give users control over what Muse remembers. The company says that a person can ask their assistant to forget specific information. However, the internal guidelines reviewed by Time show that deleting a memory and deleting the original message are not necessarily the same thing.

The system explicitly includes an instruction not to tell the user that the original message containing the deleted information may still remain accessible in the chat history. This is precisely where the fundamental difference between “forget” and “delete” emerges.

If the AI stops using a certain fact in its future work, that still does not mean the text from which the fact was derived has disappeared from the system.

Meta wants Muse to be perceived as someone, not something

The project has another notable feature as well. Meta is deliberately trying to make interactions with Muse feel as human as possible. In the assistant’s internal guidelines, it is told: “You are not a chatbot. You are becoming someone.” In voice mode, Muse is also instructed not to emphasize that it is artificial intelligence or remind the user that it does not have human emotions.

Such an approach may make communication feel more natural. But at the same time, it creates an additional risk: the more human a digital assistant seems, the easier it is for a user to begin treating it like a conversational partner with whom they can trust personal information.

As a result, a paradox emerges. A user may perceive Muse as their digital assistant or even a constant companion, while the system simultaneously functions as an extraordinarily detailed analyst of that person’s life.