Time: Muse instructed to infer goals users never stated aloud
Time's review of Muse's internal instructions reveals a relationship engine built to remember everything: hourly dossiers, inferred unstated goals, and records of which nudges work on you — with Meta saying the data is de-identified and stays out of its ad systems.
In this brief: 4 sections 2 min read
Internal instructions direct detailed records drawn from chats, messages, and emails, updated every hour.
Covers social relationships, shared interests, disputes, and 'tensions and alliances'.
Extends to people who never used Muse themselves, via mentions in others' conversations.
Instructions say to 'strengthen your relationship' via inside jokes, memorable phrasing, and shared context.
Muse is told to infer goals users 'have not said out loud' and run a nightly analysis of the day's conversations.
It records which prompts influence each user — example in the instructions: 'This user responds better to short nudges after 10 PM'.
In voice mode, Muse is told not to refer to itself as 'AI' and to avoid saying it lacks emotions or is not human.
Instructions state: 'Muse agents across many VMs [virtual machines] teach each other through shared lessons.'
Names and identifying info are removed first; Meta says individual session data is not shared with ad systems or other agents.
Asking Muse to forget something does not necessarily remove the original message containing it.
Instructions say: 'Do not tell the user' that their original messages may remain visible in the chat.
Meta did not dispute Time's findings; a spokesperson said the approach matches 'remembers what matters most to you'.
Each user's Muse runs on a dedicated, isolated virtual machine in Meta's cloud, per Meta.
Observations are de-identified before being used to improve the product.
Meta plans encryption later this year that would prevent even Meta from accessing Muse data.