Google's Pet Memory names your pets, then calls all three cats Smokey

Google's Pet Memory, part of Gemini for Home, is sold as the step past generic animal alerts: it promises to teach Google Home who your pets are so connected Nest cameras can report which animal they saw and adapt the home around it ^[1]. In two weeks of testing, The Verge's smart home reviewer reported that after she described all three of her cats to Google Home, the system insisted every one of them was Smokey, the first cat she had entered ^[2][3].

The price of that is not incidental. Pet Memory sits behind the top-tier Google Home Advanced Plan at $20 a month ^[5], or $240 a year ^[7], and it currently runs only on Nest cameras, and among those only indoor Nest Cams ^[6]. That gating alone removed one of the reviewer's three intended uses before accuracy entered the picture: an outdoor-facing digital pet door, where a camera tells you which cat is asking to come in and filters out the neighbour's cat or a chicken at the back door ^[8][9].

The remaining two uses were an indoor headcount before dark and per-cat automated feeding ^[9], and both are identity-gated. She placed indoor Nest Cams in strategic spots to check which cats were inside ^[15]; asking a Nest Hub when Smokey was last seen returned video of the new kitten ^[4]. A recognition system that answers confidently with the wrong name is worse for an operator than one that says "animal detected", because the wrong name is actionable. It ends the check rather than starting one. Her conclusion, per The Verge, is that pet-focused notifications and automations are mostly useless because the system cannot tell Boone from Smokey ^[10].

That matters more than a single household's cats, because the whole justification for the feature is notification triage. The reviewer's starting problem was that her cameras send so many alerts she misses the important ones ^[14], on a property that includes free-ranging chickens, a counter-surfing Wirehaired Pointing Griffon, two indoor/outdoor cats and a kitten ^[13]. Named recognition is supposed to be the filter. A filter that labels three animals with one name does not reduce the alert volume, it just adds a false specificity to the same volume.

The failure is also a useful reminder of how different classification and identification are as engineering problems. Nest has shipped coarse classes for years: the 2017 Nest Cam IQ separated people from pets, animal detection followed, and more recently Google added AI text descriptions of what the camera sees ^[11]. Ring, Wyze and Apple Home via HomeKit Secure Video have added similar AI descriptions that often name the type of animal, not always correctly, and Google Home is the first to personalise it ^[12]. Class detection tolerates ambiguity, because "an animal" is still true. Naming does not, and the tolerance for error a user extends to "cat" collapses when the product asserts "Smokey".

Watch whether Google's enrollment moves from text descriptions to reference images, and whether the assistant is willing to return an "unsure" instead of a guess, since a confidence threshold is the cheapest fix available here. Watch whether Pet Memory reaches doorbells and outdoor cameras, which is where the pet-door use case actually lives ^[6]. And watch how long identity-based alerts stay behind the $20 Advanced tier at this hit rate ^[5][10], because paid features are the ones users benchmark hardest.