Internal instructions examined by WIRED describe how Meta’s Muse assistant can turn information about a user’s personal contacts into an organised relationship memory. Rather than merely remembering a preference mentioned in a conversation, the system’s documentation outlines individual pages that accumulate context about family, friends, partners and colleagues. The findings make the assistant’s approach to personalisation more concrete, while raising questions about how it interprets the people around a user.
The report is based on files obtained by independent AI safety and security researcher Karan Joshi. He used Muse’s ordinary chat interface to ask it to copy and share its own software files, then provided the findings to WIRED. Meta maintains that the files were intended to be accessible for transparency. Their availability should therefore not automatically be described as a security breach.
Relationship memory goes beyond a list of contacts
The instructions describe an hourly process for gathering information about people in a user’s life, including collaborators and people the user follows. Muse’s memory consists of structured text files. An individual page may begin with little information and become more detailed as further evidence becomes available.
The proposed organisation covers factual background, shared history, the character of a relationship, common interests, unresolved matters and possible ways to strengthen the connection. Examples in the documentation include where someone lives or works, significant dates, recurring discussions and remembered experiences. Suggestions could range from revisiting something a friend mentioned to remembering an anniversary or arranging an appropriate outing.
That structure matters because it combines information retrieval with interpretation. Remembering a birthday is different from assessing how close two people are or what their relationship currently needs. Both appear within the documented approach, but they require different judgments about the available evidence.
The instructions explicitly discourage invented details and favour leaving a page incomplete over filling it with unsupported information. That is an important constraint in the design. It is not, by itself, an independent demonstration that the assistant always follows it. WIRED’s account reveals operating instructions, not a comprehensive audit of every user’s stored relationship pages.
Meta describes isolation, deletion and action controls
WIRED reports that each Muse user has a dedicated virtual machine holding personal data and context, inaccessible to other agents. Users can erase memories or disconnect external services. These controls address where context is held and whether connections and accumulated information can be removed.
Meta also says the assistant seeks human confirmation before completing actions such as sending an email or making a purchase. An audit log lets users review its activity and planned actions. Those are controls over what the agent does with its context, rather than an explanation of every inference it makes while building that context.
Meta spokesperson Daniel Roberts told WIRED that useful agents need information about both users and the people they interact with. He said Muse gathers this from public information and material users choose to share. His examples linked remembered context to practical help: recognising an invoice sender as a previously hired plumber or recalling a spouse’s preferred flowers.
This explanation presents relationship memory as a mechanism for relevant assistance. It does not remove the distinction between storing a detail a user supplies and drawing a conclusion about another person from several details.
More context brings questions about inference
Joshi interpreted the documentation as an effort to understand real relationships and become familiar with a user in a friend-like way. His reaction is a researcher’s assessment of the files, not proof of a particular unauthorised data collection event.
Carissa Véliz, an associate professor at Oxford’s Institute for Ethics in AI, raised a broader concern in WIRED’s report: people reveal more to AI systems than those systems reveal back. Information can also be inferred or assembled from multiple sources. Accurate and inaccurate conclusions create different concerns, making visibility into an assistant’s understanding important alongside access controls.
Miranda Bogen, who directs the Center for Democracy and Technology’s AI Governance Lab, said Muse appears more focused on relationships and personal contacts than rival systems. She acknowledged the transparency and editing capabilities offered by memory tools, including Muse, while noting that assistants encourage users to supply increasingly broad personal information.
The new findings expose a specific design choice: organising assistance around an evolving account of relationships. Meta describes controls for inspecting activity, removing memories and confirming actions. The remaining question is how clearly users can understand the interpretations behind the help they receive.
