Meta’s new AI agent Muse is designed to know far more than what users explicitly tell it about themselves. Internal instructions examined by security researchers show the assistant can create structured profiles for the people surrounding a user, gradually recording details about friends, relatives, partners and colleagues.
The feature is intended to make Muse more useful as a personal assistant. But it also illustrates how the rise of persistent AI memory is changing the privacy equation: information shared with an assistant may increasingly describe not only the person using it, but everyone around them.
A Page for Every Person in Your Life
Independent security researcher Karan Joshi extracted Muse’s internal instructions through its normal chat interface and shared the findings with WIRED. Meta says these files were intentionally accessible as part of its transparency efforts.
Among the instructions is a process for creating what amounts to a personal page for each important person in a user’s life. Those profiles can cover family members, partners, friends, coworkers, collaborators and even people the user follows.
A profile might initially contain very little information, but Muse can expand it as more context becomes available. Suggested sections include facts, relationship history, shared interests and unresolved conversations. The system can also remember locations, occupations, birthdays, anniversaries and events such as trips, arguments or personal milestones.
Muse is instructed not to invent missing information, but it can draw conclusions from the evidence available to it. That distinction becomes increasingly important as AI assistants gain access to emails, messages, calendars, financial information and other personal data.
AI Memory Is Becoming Relationship Memory
Memory has become a major feature across the AI industry because persistent context makes assistants more useful. Instead of repeatedly explaining who someone is or what happened previously, users can allow an assistant to retain that context.
Muse takes the concept further by emphasizing relationships. It can track how close two people appear to be, what they have in common and what might strengthen their relationship. The system could then remind someone to call a relative, revisit something a friend mentioned or suggest an activity based on another person’s interests.
Meta argues that this context is necessary for an effective personal agent. A spokesperson told WIRED that Muse uses public information and information users choose to share so it can understand, for example, that someone appearing in an invoice is the same plumber previously hired by the user.
The privacy complication is that the plumber, friend or family member being profiled may never have interacted with Muse themselves.
The Privacy Boundary Extends Beyond the User
AI assistants traditionally frame privacy around consent from the account owner: users decide whether to connect their email, calendar or other services. Persistent relationship modeling complicates that assumption because one person’s data can reveal significant information about another.
A conversation might reveal where someone lives, their financial situation, their relationships or events in their personal life. Over time, an AI system capable of combining those fragments could construct a surprisingly detailed profile.
Oxford Institute for Ethics in AI associate professor Carissa Véliz told WIRED that the concern extends beyond information people deliberately provide. AI systems can also infer information and combine details from different sources, with those conclusions potentially being either accurate or incorrect.
That makes AI memory qualitatively different from simply storing previous conversations. The assistant is organizing information into a model of the user’s social world.
Personal Agents Need Extraordinary Levels of Trust
Meta has implemented safeguards around Muse. Each user receives a dedicated virtual machine for storing context, which the company says is isolated from other users and agents. Users can erase memories or disconnect services, while an audit log provides visibility into the agent’s actions and future plans.
Muse is also designed to request confirmation before sensitive actions such as sending emails or completing purchases.
Those protections address important security risks, but they do not eliminate the broader privacy question created by increasingly capable personal agents. To become genuinely useful, assistants are being encouraged to understand emails, calendars, finances, health information and relationships simultaneously.
That produces an unusual trade-off. The better an agent understands someone’s life, the more useful it can become — but the same context also makes the system an extraordinarily concentrated repository of personal information.
Muse makes that tension unusually visible. The future of AI assistants may not simply involve models that remember you. They may increasingly build structured representations of everyone who matters to you, including people who never chose to use the AI at all.