What sits in the Memory file?
WIRED's Reece Rogers spent several days with Muse. Sensor Tower put first-week downloads above 900,000. Muse stores interactions and preferences in a Memory document opened by tapping the avatar at the top centre; the in-app text calls it durable facts, preferences and commitments. A chat request can wipe the data and the file can be edited by hand. In Rogers's test Meta did not offer a toggle that turns the feature off altogether.[1]
Under Ideas, Muse offered, the next day, to wire a savings tracker to a real bank balance, to scan the inbox, to estimate calories from meal photos and to be told when a passport and licence expire. Rogers also wrote that a Stripe prompt asked for a card for a bakery order. WhatsApp and Instagram supply onboarding and cross-promotion. Until those sources are linked, Memory may hold only the chat; when they are linked, the same file can carry context from those accounts.[1]
Does the training default sit at the same door as Memory?
Users are opted in by default to having Muse interactions used for model training. Meta says that data is sanitized; Rogers wrote that it is unclear how the process works. Collection may include what Muse uses as context from a connected inbox or bank account. Under Data controls the Help improve our AI models toggle can be switched off. There is no equivalent master switch for Memory itself.[1]
Meta spokesperson Emil Vazquez told WIRED that Muse was built for everyone with built-in protections and user controls, and that any suggestion it was not built that way from the beginning is ludicrous. In the chain Rogers followed, control sits in the training toggle and in line-by-line edits of the Memory file, not in a switch that shuts the feature. Rory Mir of the Electronic Frontier Foundation said the chat window writes to the hosting company's servers. If a user never links a bank or inbox, the training default may carry only chat text; the prompts Rogers saw ask for exactly those links.[1]