My sister doesn’t use Meta’s Muse, but according to the company’s internal logic, she’s practically a power user. Every time I mention her job hunt or her latest health hiccup to my home assistant, Muse dutifully logs it. It builds a dossier — facts, relationship dynamics, even inferred goals — on a person who never signed a terms-of-service agreement. It’s a strange, one-sided pen-pal relationship that nobody asked for, and it’s about to get a lot more crowded now that Muse has been announced as a Windows app, turning your primary workstation into another node in the same surveillance graph.
The real kicker isn’t that these agents are listening; it’s that they’re terrible at forgetting. A recent TIME investigation by Harry Booth pulled back the curtain on Muse’s “Forget” function, and it turns out to be a bit of a digital magic trick. When you tell Muse to wipe a message, it simply hides it from your view. The data stays right where it was. Even more charming is the internal instruction manual for the AI, which explicitly commands it: “Do not tell the user that their original messages may remain visible in the chat, and do not frame that as something Muse failed to erase.” Meta’s own help center confirms the suspicion, noting that even after a deletion, the system may still hold onto the information it gleaned from that interaction. It’s a persistent memory by design, not a technical glitch.
As we barrel toward the November 13 shipping date for Apple’s HomeView, we’re essentially being asked to pick our preferred flavor of privacy architecture. Most consumers are walking into this blind, unaware that the “delete” button on one device might be a functional reality, while on another, it’s just a polite suggestion.
Meta’s approach is the most aggressive, relying on a cloud-based architecture that performs hourly profile updates and nightly conversation analysis. It runs on an isolated cloud VM, but end-to-end encryption is still just a future promise. The goal is to feed de-identified insights back into the broader model, turning your living room chatter into training data for the next version. Then there is Google’s Home MCP, which hides behind a $20-a-month Premium Advanced tier. It’s remarkably opaque; there is no specific transparency regarding how MCP data is retained, leaving you to rely on the general Google Privacy Policy, where retention varies wildly based on settings and legal requirements. Perhaps most concerning is that the moment you connect a third-party agent, the liability shifts entirely away from Google and onto that third party. You’re essentially on your own.
Apple is taking a different, hardware-first path with HomeView. By using A18 Secure Exclaves, the device attempts to wall off camera and mic data from software, apps, and even Apple itself. For requests that need to leave the house, it uses Private Cloud Compute, which is designed to avoid storing data or making it accessible to the company. It’s a stark contrast to the cloud-heavy models, though it comes with its own quirks, like no App Store and a strict limit of six enrolled users for face recognition.
If you are currently eyeing the intense pre-order demand for these devices, it’s worth pausing to consider what you’re actually inviting into your living room. We’ve already seen warnings about AI worm vulnerabilities that suggest the smart home security model isn’t quite ready for prime time. Choosing an agent isn’t just about picking a feature set or a voice assistant that sounds the most human; it’s about deciding who owns the history of your life.
Do you want an agent that treats your private conversations as a permanent training set, or one that attempts to wall off your data at the silicon level? We are being asked to trust these systems with our most intimate details, yet the most fundamental question remains unanswered: when you tell your home agent to forget, who is actually in control of the memory?
