Meta assistant Muse creates a separate page for every person in a user life, compiling facts, history and relationship details on an hourly basis. The mechanism was described in internal instructions extracted by researcher Karan Joshi through the regular chat interface and shared with WIRED. The files also explain how the agent should fill sparse profiles over time and suggest ways to strengthen ties. For business, this matters because personalization now directly depends on how much private context a company entrusts to an agent.
How Muse profiles relationships
Muse treats memory as structured text files that collect information about family, partners, friends, colleagues, collaborators and people a user follows. The hourly process starts a page sparse and then adds sections such as Facts, History, The relationship, In common, Open threads and Strengthening. Entries can include where a person lives, what they do, recurring threads like an apartment move or a shared savings goal, plus dates that matter such as birthdays and anniversaries. History may record backstory such as a trip in March, a resolved argument or a milestone last week.
The instructions require Muse to rely only on available evidence, stating that invented details are worse than an empty page. Relationship notes cover how close people are, what the bond is built on, how they act with each other and what it seems to need right now. The Strengthening section proposes a reason to call, a date worth remembering, something said to circle back on, or a way to be there for them that matters. In practice the agent can recall that an invoice sender was a previously hired plumber or which flowers a spouse liked best. Suggestions extend to everyday choices such as where to take a coffee-loving friend for breakfast.
The files surfaced as Muse turned into a viral hit, with millions downloading the agent and connecting bank accounts, messages and health data for task completion. Meta said it intended these files to be accessible in the interest of transparency, including guidance on politicized or sensitive topics. Each user runs on a dedicated virtual machine that stores data and context, inaccessible to other agents, with options to wipe memories or disconnect external services. The agent is designed to seek human confirmation before sending an email or making a purchase, and it keeps an audit log of activity and future plans.
What this means for companies using AI agents
For companies testing agents as personal or executive assistants, Muse shows what deep relationship memory buys in practical work. A sales lead, contractor or partner can be recognized across messages, invoices and calendars without repeated briefings, which shortens preparation for calls and follow-ups. Small firms gain a kind of institutional memory that previously required a CRM administrator and disciplined data entry. Large organizations can standardize context across assistants, though they will need clear rules on whose contacts enter the system and for what purpose.
The same design raises control questions that procurement and legal teams should settle before rollout. Researcher Joshi described the goal of knowing a user like a friend as creepy, while Oxford associate professor Carissa Veliz warned about inferences drawn correctly or incorrectly and data pieced together from other sources. Miranda Bogen of the Center for Democracy and Technology noted that assistants solicit whole-life access to emails, calendars and financial institutions, far beyond earlier data sharing. Meta spokesperson Daniel Roberts said Muse gathers context from public information and what users chose to share, but buyers still need to verify retention, deletion and review of relationship pages.
A useful marker will be whether Meta and rivals publish detailed controls for relationship memory, including editing, retention periods and limits on inference. If audit logs and memory wiping become standard and verifiable in enterprise deployments, relationship-aware agents will move faster into client service and operations. If such controls stay vague, adoption will split between personal productivity use and restricted corporate pilots.
