— a companion that remembers you
Published 19 June 2026
We're building a small, fully on-device companion — Risa. It can't be a know-it-all assistant. It's trying to be something rarer: a warm, expressive presence that remembers you, entirely on your own phone.
Most AI today races to be the smartest thing in the room. We're chasing a different feeling. Our name for it is a little pun that only works out loud: AI = 爱 — in Mandarin, "AI" sounds like ài, the word for love. An AI that cares and remembers, not one that merely knows.
The bet: small model, big heart
Risa runs on a small language model — only a couple of billion parameters — living entirely on your device. Let's be honest about what that means: a model this size is not a fountain of world knowledge. It hasn't read the whole internet, it can get facts wrong, and it won't out-reason the giant models in a data centre. If you want an oracle, this isn't it.
But "smart" was never the point. Our wager — and the entire design of the product — is this: a small model can't be a great assistant, but it can be a great companion. Being a good companion doesn't require genius. It requires being consistent, being warm, being present, and above all remembering you. Those are engineering problems, not IQ problems — and they're the ones we've chosen to solve.
The one promise that shapes everything: it all stays on your device
A companion only works if you can be honest with it, and you can only be honest with something that keeps your secrets. So Risa is built the same way the rest of SpecEnvoy is: everything runs locally. Your conversations and the things it remembers about you live on your phone, not on our servers — because there are no servers in the loop. Nothing is uploaded, nothing trains anyone's AI. The only time the network is touched is the one-time download of the model itself.
That isn't a feature we bolted on; it's the constraint we designed around, and it's why the model has to be small enough to live in your pocket in the first place.
It remembers you — and that's the hard part
Anyone can make a chatbot reply. The thing that turns a reply into a relationship is memory that lasts — so that weeks later it gently asks how your mum's appointment went, or remembers that you hate mornings and love your dog. That's the product. The face and the chat are just the window; memory is the house.
Here's the genuinely interesting research problem hiding inside that. The information a small companion learns about you arrives noisy and uncertain — a small model can mishear an intention, and (once we add voice) speech-to-text can mishear a word. If you naively wrote all of that down as fact, your companion would slowly fill up with confident nonsense.
So we don't. Our approach is best described as memory engineering for low-confidence inputs: every detail Risa learns is treated as provisional — tagged with where it came from and how sure we are — and it only firms up into something Risa will rely on as it's corroborated over time. And because it's your memory, you can always see it and correct it. The guiding principle, the part we're proud of: reliability is engineered around the model, not demanded of it.
A face with feeling
A companion needs a face you actually want to look at. Risa's is a small, glowing, screen-style face — and it's far more expressive than its simplicity suggests.
We didn't reinvent expression from scratch; we leaned on a century of people who got there first. Risa's movement follows the classic principles of character animation — the anticipation, follow-through, and arcs that make motion read as alive rather than mechanical (as the old animators put it, only robots move in perfectly straight lines). And its emotions follow a dimensional model of feeling: every mood has both a kind and an intensity, so the same warmth can be a quiet smile or an overflowing, can't-contain-it beam. From a deliberately small, low-overhead face, that yields a surprising amount of range — calm and giddy, tender and teary, curious, sleepy, bashful — plus the little unprompted gestures and idle life that make it feel like someone's home behind the screen.
You can meet her live — the renderer now runs right on the page, and she'll talk to you.
It runs on the phone you already have
None of this needs a server or a subscription. Risa is designed to run comfortably on a mid-range 2024 phone with 8 GB of memory — privately, offline, yours. The cost of a companion shouldn't be your data.