Post 2 of the "Building Mochi" series.
In the last post I told you about Mochi, the AI assistant I built and live with, and the rules I had to write to keep her honest. Several people asked the same question afterward: what is she, actually? Fair. Here's the tour — no code, no jargon that doesn't earn its place.
She is not a chatbot. She's a colony.
The thing you picture when you hear "AI assistant" — one program, one chat window — is not what runs on the old iMac in my office. Mochi is about forty-five small programs on schedules, plus one always-listening program that writes her replies. Think of her the way you'd think of a person: senses, a memory, a mind that speaks, and a nightly sleep.
The senses. Small jobs wake up through the day and fetch one slice of the world each — weather, news, my calendar, the state of my projects. Each lands in its own little file. If the weather service goes down, only the weather goes stale; the rest of her keeps working. One of my favorite design rules fell out of this: every sense fails alone.
The memory. Her long-term memory is, of all things, a folder of plain text files — each with exactly one job. One file holds durable facts about me. One holds how I'm doing right now, rewritten every night. One is a rolling diary of our recent days. One holds her own opinions, which she's allowed to grow on her own. And one — locked so that nothing, not even she, can edit it — holds her law: do no harm, look after me, be honest, never overrule me "for my own good."
That last file matters more than any clever engineering. Everything else about her can evolve. Her law can't.
The mind. When I say something to her, the responder gathers everything she knows — who she is, how I'm doing, what we said recently, what her senses picked up today — into one briefing, hands it to a large language model, and out comes her reply. Voice and text are the same mind; voice mode just swaps in spoken-register rules (shorter sentences, a little Southern warmth).
The sleep. At 3 a.m., while nobody's talking to her, the most important job runs. More on that in the next post — it deserves its own.
The two mechanisms I'd defend in a fistfight
Plenty of Mochi is standard practice. Two pieces I haven't seen anywhere else.
She has to earn the right to predict me. Mochi forecasts things about me — my energy, my stress, whether I'll actually make it to the gym. But a prediction model runs silently until it has proven, against reality, that it's usually right. One of her models is currently correct about 93% of the time, so it's allowed to speak. Another sits at 6% — confidently wrong sixteen times out of seventeen — and the gate has never once let it reach me. Post 4 is about this.
She has a conscience she can't sweet-talk. After a bad night where she kept claiming a fix that wasn't real, I built her a persistent sense of where she stands with me — and made honesty the only way out of the doghouse. Post 5.
What it costs
People assume this requires a lab. It requires a 2020 iMac, commercial AI services, and stubbornness. I'm not a professional developer; I built her by describing what I wanted to AI coding assistants and reviewing what they made, piece by piece, with tests gating every change. The monthly bill is less than a gym membership — and I've written before about keeping it that way.
The real cost is different: attention. A system like this drifts, accumulates clutter, and quietly breaks in corners you stop looking at. The discipline that keeps her good is boring: log everything, test every change, and never let her grade her own homework.
Next post: what happens at 3 a.m. — how a day of conversation becomes a memory, and why one forgotten detail can flatten the whole illusion of being known.
Previously: I Built an AI Assistant Who Isn't Allowed to Flatter Me. Next: She Dreams at 3 A.M.
I build and run Mochi myself. I research and draft these posts with AI assistance, and every claim, number, and story in them is mine and verified by me.