Muse’s AI Memory Builds Detailed Profiles of Users’ Relationships

Muse’s memory system is designed to build a page for each person in a user’s life, drawing together details that could help the assistant make suggestions and follow up on past conversations. The feature points to a broader shift in personal AI: assistants that connect to email, calendars and other services can assemble a picture of someone’s relationships, not just remember isolated preferences.

Instructions extracted from Muse describe an hourly process for compiling information about family members, partners, friends, colleagues, collaborators and people a user follows. The pages can begin with little information and grow as the system encounters more context. Their possible sections include “Facts,” “History,” “The relationship,” “In common,” “Open threads” and “Strengthening.”

What Muse’s relationship pages may contain

The instructions describe entries such as where a person lives and what they do, along with recurring topics: an apartment move, a shared savings goal, or something the user meant to revisit. Muse may also record dates such as birthdays and anniversaries, and add events to a person’s history, including a trip, a resolved argument or a recent milestone.

The “Strengthening” section is meant to turn those details into possible next steps. It might suggest calling someone, remembering a meaningful date, returning to a topic they raised, or offering help in a way that suits the situation. A coffee preference could inform a breakfast suggestion; a note about a plumber could help the assistant connect an invoice to someone previously hired for work at home.

These are intended uses described in Muse’s instructions and by Meta. They don’t establish that every suggested profile will be accurate or complete. The documentation says pages may remain sparse and directs Muse to rely on available evidence. It also says invented details are worse than leaving a page blank. That instruction matters because a system that infers personal connections can produce a confident but mistaken account if it treats a guess as a fact.

From saved details to inferred context

AI assistants commonly offer memory features to personalize answers and tasks; users have also sought relationship guidance from chatbots. Muse’s design appears to put more emphasis on organizing information about other people in a user’s life, according to Miranda Bogen, director of the AI Governance Lab at the Center for Democracy and Technology.

Karan Joshi, an independent AI safety and security researcher, extracted Muse’s instructions and system prompts through its chat interface by asking the assistant to share its own software files. He shared the findings with WIRED. Meta has said the files were accessible by design as a transparency measure. The extracted material offers a view into how Muse is instructed to handle user requests and organize context.

Joshi said the instructions suggested that Muse aims to understand users through their relationships with real people. He described that approach as unsettling. His reaction gets at the central tension in this design: a memory system can make an assistant more useful by connecting details across interactions, while also creating a more extensive record of a user’s social life.

Carissa Véliz, an associate professor at Oxford’s Institute for Ethics in AI, says the information involved goes beyond what people deliberately provide. Systems may also infer details, correctly or incorrectly, and combine them with information available from other sources. An inaccurate inference could affect the advice or reminders a user receives; an accurate one could still reveal more than the user expected the assistant to retain.

Controls and the scale of access

Meta says each Muse user has a dedicated virtual machine to store data and context, inaccessible to other users and agents. Users can erase memories or disconnect external services. Meta also says Muse seeks confirmation before actions such as sending an email or making a purchase, and provides an audit log showing the agent’s activity and planned actions.

Those controls address different parts of the experience: users can inspect activity, remove stored context and interrupt connections to outside services. They don’t, by themselves, answer how useful or accurate a relationship profile will be, or how readily someone will notice an incorrect detail. The practical value of the feature depends on the quality of its records and on whether users can understand and correct what Muse has retained.

Meta spokesperson Daniel Roberts told WIRED that Muse gathers information from public sources and from details users choose to share. He said the context can help the assistant connect a person to an earlier interaction, such as recognizing that someone who sent an invoice is a plumber the user hired before, or recalling which flowers a spouse liked.

Bogen points to a wider design pressure: assistants invite people to connect more of their lives to make the service more helpful. Email, calendars and financial accounts can each add useful context, but together they expose a much broader range of personal activity. Tools that offer transparency and editing still put the burden on users to review what has accumulated.

Muse’s relationship pages make that tradeoff concrete. The assistant may remember the people and open threads that would otherwise slip through the cracks, but the same capability asks users to trust its collection and interpretation of intimate details. How clearly Muse shows the basis for each entry—and how easily users can fix a mistaken inference—will shape whether this kind of memory feels like useful assistance or an unwanted profile.