Google's Pet Memory forgot who my cats are

Google launched its Pet Memory feature for Nest cameras in June 2026, promising to turn generic "a dog is in the kitchen" alerts into personalized "Fido is in the kitchen" notifications — but a two-week real-world test published by The Verge on August 18 found the system unable to tell three cats apart, raising a pointed question about where Gemini's contextual recall actually breaks down.

Background You Need

Smart home security cameras have spent years getting better at detecting motion, classifying objects, and filtering out false positives — a squirrel versus a package, a person versus a shadow. What they have never meaningfully done is distinguish which person or which pet triggered an alert. For households with multiple animals, that gap matters: a notification that says "your cat is outside" is far less useful than one that tells you whether it's the one who isn't supposed to be out.

Google has been layering Gemini AI across its Home platform in 2026 as part of a broader push to make the smart home feel responsive rather than mechanical. The "Ask Home" voice feature — which lets users query their home's camera feeds and sensor data through a natural-language interface — has been central to that push. As of early June 2026, Google expanded Ask Home to more than 20 countries where Gemini for Home is supported, including Australia, Canada, France, Japan, the UK, and the US, while simultaneously bringing Germany into early access. Pet Memory was announced alongside that expansion as its headline consumer feature, reported by 9to5Google's Ben Schoon on June 2.

The feature sits behind Google Home Premium Advanced, Google's paid subscription tier — an important caveat that BigGo's coverage of the June rollout flagged explicitly. The idea is straightforward: through Ask Home, users describe their pets by name and type, and Gemini is supposed to use that stored description to personalize future Nest camera notifications and answer real-time questions. Marketing copy positions it as reducing misidentifications, ensuring owners "know exactly which pet is causing a ruckus."

What's New

The August 18 review by Jennifer Pattison Tuohy at The Verge is the first extended hands-on account of Pet Memory operating in a real multi-pet household over time, and the headline result is unambiguous: after two weeks of use, the system could not reliably distinguish one cat from another.

The failure pattern is instructive because it cleaves along a specific line. Species-level detection — "there is a cat in this frame" — works. That's the baseline Nest cameras have delivered for years. What fails is the next layer: individual discrimination, determining that the cat in this frame is Boone and not Smokey. That distinction is exactly what Pet Memory is supposed to deliver, and it is exactly where the system collapses.

This failure mode is more damaging than a simple miss. In a system where memory is supposed to gate automated actions — send a notification tagged to a specific pet, trigger a specific routine — a wrong identification does double harm: the wrong automation fires, and the correct one never does. The error doesn't just fail to add value; it actively degrades the baseline experience the camera provided before Pet Memory was layered on top.

Multiple earlier sources, including BigGo's June coverage, described Pet Memory with confidence, quoting Google's own language about how it "helps reduce misidentifications." That framing aged poorly within weeks of the feature shipping broadly. The gap between how Google positioned the feature at launch and how it performed in testing is notable, though not entirely surprising — the June reports were drawing on Google's announcement language, not on sustained real-world evaluation.

Our read, drawing on both the Verge's testing and what the corroborating sources establish about the feature's architecture, is that this is a structural problem, not a configuration one. The feature asks users to describe pets in text, then expects Gemini to map those text descriptions onto camera frames in real time. For visually similar animals — three cats of different colors, let alone similar ones — that text-to-vision matching appears insufficient. Google hasn't published any technical documentation on how the individual discrimination model works, which makes it hard to assess whether this is a training data gap, a model architecture limitation, or something addressable through more user-provided input.

The Pushback

The honest counterpoint is that two weeks of testing with three cats is a narrow sample, and the feature is still in early access. Google could plausibly argue that more labeled examples, more camera angles, or continued model updates will improve individual recognition over time. Some smart home AI features have genuinely improved meaningfully in the months after launch as the underlying models were refined.

But that argument cuts both ways. A feature locked behind a paid subscription tier, marketed on the premise of telling your pets apart, should not require the user to wait out a training period of unknown length before delivering its advertised value. The 9to5Google coverage of the June announcement framed the setup process as simple — name and description, through Ask Home — with no mention that recognition accuracy might degrade in multi-pet households or require ongoing correction. If Pet Memory's viability depends on factors like coat pattern distinctiveness, camera placement, or the number of animals in the household, those constraints need to be surfaced before purchase, not discovered after two weeks of failed identifications.

There is also a broader architectural lesson here that extends beyond smart home cameras. Any system that chains stored entity memory to automated downstream actions inherits both error modes when recognition fails: the wrong action triggers, and the right one doesn't. Google's Pet Memory is a consumer product, but the same failure pattern shows up anywhere developers wire identity recognition into conditional workflows. The confidence of the recognition step needs to be an explicit gate, not an assumed pass — something the Verge's testing surfaces in concrete, domestic terms.

Sources