Every conversation with an AI starts from zero. The context you built last week is gone, and whatever it learned about your domain went to a model you don't own.
Memory doesn't compound. A tool that forgets can't get better at your problem. It stays a generalist forever, and you re-explain yourself forever.
The value flows the wrong way. The understanding you generate becomes training data for someone else's product — extracted upward, not accumulated where you stand.
Expertise stays trapped. The person who knows how the system actually runs can't hand that knowledge to anyone — it walks out the door when they do.
Not "an AI that remembers" — every product says that. Something more specific: a system designed so that perspective on a domain becomes an asset, owned where it's generated, and transferable without being given away.
A sovereign brain isn't one giant pile of everything. It's a protégé assigned to a perspective — this aquaponics system, this apiary, this curriculum — that learns under that lens and gets sharper at exactly one thing. Cross-domain search still works when you want it. One substrate, many specialists.
A small model in a grow tent holds the local view of its own project and writes to a brain you control. The most sensitive material never embeds externally at all. The architecture follows the work; the data stays where it was earned. Escalating to a bigger model is a choice, not a default surrender.
A brain that stores notes is a database. A brain that's mature enough to brief a collaborator, draft a reply, and hold context across weeks on the project's behalf is the actual deliverable. Storage is the seed. Mediation is the harvest. And a mature brain can teach a peer brain the pattern — without handing over what it knows.
The claim above is easy to make and hard to believe. So it's being built in the open. Node Zero is a cold-water aquaponics pilot in a New England basement — a 30-gallon tank, goldfish and parsley, running while the sensor stack gets hard-soldered on the bench; the instrumented 120+ gallon build moves into the garage in September. The brain behind it is already live: ~80,000 captures, growing since late March 2026.
The Live page is wired to the real API and shows exactly what the system reports — including an honest offline state until the tank sensors come online. No projections, no invented numbers.
The dataset isn't a byproduct. It's the point. The longer the brain runs, the smarter the system that runs on it gets — accumulation made literal.
The operations manual outline is public, and the full manual is being ported section by section. Replicable on purpose — that's how you prove a pattern.
Build notes, decisions, session captures — all in one queryable memory that decides what's worth surfacing tomorrow. The substrate, working at the smallest honest scale.
Every capture — a note, a decision, and soon a sensor reading — becomes a node in one growing memory. This is that memory, live — each point a capture, each line a relationship the brain drew on its own. Locked nodes stay locked: the shape crosses the wire, the contents never leave the box.
Here's my brain. What's yours look like?
Node Zero is proving the shape: a small loop → a scoped brain → memory that compounds → a documented, transferable pattern. Nothing about that shape is specific to fish. Swap the inputs and the same architecture runs a different domain. Each new node is a peer, running the same playbook somewhere else.
Cold-water pilot, basement. A 30-gallon tank — goldfish and parsley — while the sensor stack is soldered on the bench. The 120+ gallon instrumented build lands in September.
The colony is established and working; the brain that watches it is a winter build.
A FarmBot bed — built, working through track calibration. Soil as a domain — same architecture, different inputs.
A classroom, a practice, a workshop, a farm. If you care about a domain, it can have a brain.
A brain is only as good as what reaches it. The same architecture takes input from wherever you already are — the point isn't the sensor, it's that heterogeneous streams land in one memory and get reasoned over together.
pH, temp, dissolved oxygen, soil moisture — a Pi on the loop, writing to the brain. The Node Zero pattern; the stack is on the bench now.
An AR/wearable stream joining the same loop — a vivid, sensor-rich feed that proves heterogeneous inputs reason together, not just side by side.
A note, a decision, a photo — typed or dictated from where you work. Lands at home, typed and deduplicated on the way in.
Conclusions worth keeping get captured straight from a working session into the brain — so the next conversation doesn't start from zero.
Field notes, build logs, and the occasional dinner — written while the solder is still warm. Don't take the homepage's word for any of it; read the notes.
First real measurement of the margin-conservation law on the sovereign corpus: the gap between the right memory and the best wrong one closes from 0.090 to 0.011 as the corpus grows.
The day cortex_watch went live — and routed its own evidence into the right folder.
First in a series. Half a London broil from earlier in the week, a pound of ground veal, a Dutch oven, and a Claude tab open the whole time. Notes on what worked.
Each node is sovereign on its own and stronger in a network. The brains trade the pattern, never the data. That's the whole bet, drawn small enough to be honest about.