TIME Finds Meta’s Muse AI Agent Builds Detailed User Dossiers

Meta’s Muse AI agent is attracting attention not only for its ability to perform tasks autonomously, but also for how extensively it can learn about the people using it. According to a TIME analysis of Muse’s internal instructions and files, the system creates continuously updated dossiers about its users and people mentioned in their conversations, messages and emails. The reports can include information about relationships, interests, disagreements and social connections, effectively creating a detailed map of a user’s personal world.

The investigation found that Muse can update these dossiers frequently as it processes new information. The system is designed to use what it learns to provide more personalized assistance and determine which suggestions or “nudges” may be useful to a particular person. Meta says users remain in control of their Muse and can decide which applications and services the agent can access. The company also says information used to improve the broader product is de-identified rather than directly shared between individual users’ virtual machines.

The findings are particularly significant because Muse is designed to have much deeper access to a person’s digital life than a conventional chatbot. Meta’s agent can operate inside a dedicated cloud computer, browse websites, send emails, fill out forms, make purchases and continue working after a user closes the application. Meta says Muse uses a separate Sentinel security system to approve internet activity and asks for permission before sensitive actions such as sending an email or completing a purchase.

However, TIME’s investigation raises questions about what happens when users ask Muse to forget information. According to the report, Muse’s internal instructions indicate that asking the agent to forget something does not necessarily remove the original conversation in which that information appeared. TIME also reported that the instructions tell Muse not to emphasize this limitation when responding to users. Meta, meanwhile, has said it plans to introduce Muse Confidential VM later in 2026, allowing users to encrypt their entire virtual machine with a key controlled only by them.

The controversy highlights the larger challenge facing the emerging personal-AI industry: powerful agents need access to increasingly sensitive information in order to become genuinely useful. Muse is already positioned by Meta as an assistant capable of managing parts of a user’s online life, while competitors such as OpenAI’s Dots are pursuing similar always-on capabilities. As these systems become more autonomous, the central question may no longer be simply what AI can do, but how much of a person’s private life users are comfortable allowing an AI system to understand and remember.