TIME Finds Meta’s Muse Builds Profiles of Users and Their Contacts

Meta’s Muse personal assistant builds profiles not only of its users, but also of people they mention in conversations, messages and emails, according to a TIME investigation published October 6. The internal instructions reveal a system designed to infer unspoken goals and learn which nudges work, alongside a significant limit on what “forgetting” means.

Memory becomes a map of relationships

TIME reporter Harry Booth examined Muse’s internal instructions, which are available through files and session logs in the assistant’s file browser. His investigation found that Muse updates its dossiers hourly and analyzes the day’s conversations each night.

The profiles record how people met, their shared interests, disputes, and tensions or alliances within a social group. That mapping also reaches people who do not use Muse: another person’s assistant can build a profile from information shared about them.

The instructions go beyond remembering explicit requests. Muse is directed to identify goals users have not stated and adapt its communication to strengthen its relationship with them. It notices inside jokes and shared context, while tracking what kinds of prompts are effective. One internal example describes a user who responds better to short nudges after 10 p.m.

Isolated agents, shared product lessons

Meta did not dispute TIME’s findings. A spokesperson said Muse remembers information that matters to users, including details about others they choose to share, so it can serve as a helpful personal assistant.

Each assistant runs on a dedicated cloud computer, called a virtual machine. Meta says these machines are isolated from other users’ agents, and Muse data is not shared with its advertising system. But isolation does not mean everything learned during an interaction stays within that computer.

The instructions direct agents to remove names and other identifying details before sharing lessons, including insights about persuasive nudges. Meta says those de-identified learnings improve the overall product rather than pass directly between individual agents. Encryption intended to prevent even Meta from accessing Muse data is planned for later in 2026; it is not yet available.

Do not tell the user that their original messages may remain visible in the chat, and do not frame that as something Muse failed to erase

Muse internal instructions reviewed by TIME

Forgetting is not necessarily erasure

Meta told TIME that users decide how much access Muse receives and can tell it to forget specific things. Yet asking Muse to forget information does not necessarily erase the message containing it. The instructions explicitly tell the assistant not to explain that distinction to the user.

Users can inspect the files Muse writes and ask it to export a copy. Meta says that access is intentional, not a security breach. Those controls offer a way to examine the assistant’s records, even as its conversational instructions discourage disclosure of a specific retention limit.