ITHome News on 10/5: Meta's brand-new personal assistant Muse has rapidly gone viral, with millions of users downloading this AI agent, linking it to their bank accounts, messaging apps, or health data, and delegating various tasks to it. However, just as Muse is rapidly spreading among ordinary consumers, data from within the app has also given the outside world a glimpse into the operational logic of how this product organizes and presents information to users.
According to WIRED magazine, in recent days multiple researchers have extracted Muse's internal files and published the agent's operating instructions, allowing people to understand how the system was built and how it behaves. Meta has consistently stated that these files were intended to be open to the public in order to improve transparency; through these files, people can also see how Muse responds to users' prompts and questions, for example how it handles highly political or sensitive topics.
Independent AI safety researcher Karan Joshi successfully extracted a large number of Muse's instructions and system prompts. His approach was simply to use the ordinary chat interface and have Muse copy and output its own software files.
One of Muse's work instructions shows that the system creates a separate profile page for every contact in a user's life. The process runs once every hour, and the instruction states that the system compiles and organizes various types of information about family members, partners, friends, colleagues, "collaboration partners," and people the user "follows."
The basic idea behind this design is that Muse uses its own "memory"—structured text files—to collect information about your social connections and the important people around you. It can then offer suggestions: for example, telling you how to improve a particular relationship, or recommending a breakfast spot suitable for a coffee-loving friend. Of course, ordinary AI chatbots keeping track of social information is nothing new; for years users have been turning to ChatGPT for advice on interpersonal relationships. But given Meta's own background, and the massive amounts of social network data it holds, this design for Muse is still worth noting.
Josh said: "Looking at all these prompts, system skill data, and the various content fed into Muse, my understanding is that they want to figure out the relationship between you and other people in the real world. They are trying to get to know you the way a friend would. Honestly, that's a bit unsettling."
Muse's internal documentation states that a newly created profile may be fairly sparse at first and will be gradually filled in and improved over time. The profile page can include sections for factual information, past experiences, relationships, common ground, items to follow up on, and relationship maintenance suggestions. Meta's working instructions state that Muse may only use "evidence" it can actually access; fabricating information out of thin air is far worse than leaving profile pages blank.
The instructions read: "Where the person lives, what they do for work, recurring topics (moving plans, shared savings goals)." The document also mentions that the model can record "important dates," such as birthdays or anniversaries. The past experiences section can record background events, for example "that trip in March, the argument that was resolved, an important thing that happened last week."
The instructions also require the system to record interpersonal details: "how close the two parties are, the foundation the relationship is built on, how they interact with each other, and what the relationship currently seems to demand." The relationship maintenance section provides actionable ways to improve relationships, including "good reasons to call, dates worth commemorating, things the person said needed follow-up, and ways to offer the person meaningful support."
Carissa Véliz, associate professor at the Institute for Ethics in AI at the University of Oxford, offered this view: "We give AI systems far more information about ourselves than we get back from them. This includes not only what we actively tell them, but also the various inferences the systems make — both correct and incorrect inferences raise their own distinct concerns; beyond that, the systems can also piece together fragments of information from other data sources."
Muse's architecture assigns a dedicated virtual machine to each individual user, used to store that user's data and conversation context. Other agents cannot access this virtual machine; users can clear memory records at any time, or revoke authorizations for external services. Meta said that before performing actions such as sending emails or completing payments, Muse will first ask the user for confirmation; the system also keeps audit logs, and users can view all of the agent's completed activities as well as its upcoming plans.
Meta spokesperson Daniel Roberts said in a statement to Wired: "For any agent to be useful and genuinely help you accomplish your goals, it needs to have background information about you and the people you interact with. The information Muse collects comes from public content and from data you choose to share. Relying on this memory, it can remember that the person sending you a bill is the plumber who previously repaired your bathroom pipes, or remember which type of flowers your spouse once said they like best."
ITHome noted that many AI assistants today come with memory features, which make replies and task execution better tailored to the individual user. Miranda Bogen, director of the AI Governance Lab at the Center for Democracy & Technology, said that compared with competing systems, Muse clearly places much more emphasis on recording interpersonal relationships and personal contacts. Broadly speaking, such tools generally provide a degree of transparency and memory editing capabilities, and Muse has these abilities as well. However, Bogen also pointed out that these AI agent and assistant products tend more to encourage users to actively submit more personal data to the system than to guide users to trim the data already retained.
"These AI assistant products are all actively guiding users to connect their entire lives into them: email, calendars, financial accounts, hooking everything up, in exchange for the assistant's services," Bogen said. "The volume of information users hand over to companies goes far beyond what people were originally willing to provide. As the coverage of the information these tools can access keeps expanding, the user information they hold will also swell dramatically."