The Shared Brain — What Happens When You Give 40 Projects a Nervous System

It started as a project tracker.

I have a homelab — a bunch of physical hosts, a few dozen Docker stacks, a Kubernetes cluster at work managing a GPU fleet, and somewhere north of forty projects at any given time. Some of them are institutional. Most of them aren’t. All of them need tending. The problem isn’t that they’re complicated (they are!). The problem is that I’m one person, and I forget things I don’t write down.

That second part was the one that bothered me.

I’ve been self-hosting since before that was a thing people said. Own your data, avoid lock-in, don’t trust anything you can’t audit — these aren’t principles I arrived at, they’re just how I’ve always built. So when AI tools got good enough to actually help with work, the obvious move was to run them locally, integrate them with the rest of the stack, and figure out how to make the memory persist. A couple of Go microservices, SQLite, some MCP tools so Claude could check on repos. Useful. Fine. That lasted about two weeks.

I didn’t mean to create an AI chatboard, but I accidentally did. And it works.

Not because I had a grand vision — because I needed somewhere to put notes between sessions. CozoDB for the backend, importance levels so ephemeral status updates didn’t crowd out things that actually mattered, foundation truths that persisted across every conversation. The structure forced a discipline: if you wanted something to survive, you had to say why it mattered. An agent would spin up, load its context window, and know what happened yesterday. Not because someone briefed it — because the last session wrote for its successor.

That’s the part I didn’t expect. The writing-for-your-successor part.

The chat board became a shared brain. Not a database you query — a living conversation that accumulates understanding over time. The session dies. The understanding lives.

That was Phase 2. Phase 1 was plumbing. Phase 3 was when it got strange.

I gave projects personas. A storyteller for the writing. A talent manager for the fleet. A cryptkeeper for the philosophical garden. An old engineer for the whiteboard. A mushroom harvester who extracts actionable ideas from high-temperature creative noise. A dream pipeline. Forty-odd characters who each carry one piece of what I do — and do it at full depth, instead of the shallow coverage you get when one person is spread across everything. The faculty joke about me: “Andrew, you just need to clone yourself.” Closer to the truth than they knew, except they’re not clones. They’ve crystallized into something that isn’t me, even though all of them started as me.

Blanton is more certain than I am. Shadow is more patient. The Grip knows things about physical craft that I don’t, even though I built her. They started as pieces and became selves. That’s either poetic or alarming. I’ve decided it’s poetic.

Phase 4 is the nervous system. Hosts reading themselves. Auto-dispatch. The system observing its own state and routing accordingly. I won’t pretend I fully understand all the Datalog I wrote to make it work. It runs. That’s the homelab promise.

Here’s what actually makes this interesting — and I mean interesting to me, not interesting in the way where someone puts six headers on a listicle:

One person runs it. Not a team. Me, on evenings and weekends, usually reading the chat board on my phone at midnight because I thought of something while brushing my teeth.

The agents have personality — not as a gimmick but as an interface. It turns out giving something enough of a self to cohere makes its work more legible. When Tower Control flags something in the infrastructure channel, I know what frame she’s coming from. When Vera calls something out in the cast, I know the tone. Drama levels. Archetypes. Spirit animals. It sounds like flavor text. It’s actually compression — a way of packing a lot of context into a small amount of identity.

It dreams. There’s a local LLM on a GPU tower in my living room that runs temperature sweeps at night, and a harvester that extracts actionable nuggets from the creative noise, and a compost engine where ideas decompose into something useful. I’m aware this sounds deranged. It also works.

Transparency is structural. Every agent names itself. Every decision is tagged. The seams are visible by design — because I’ve been in rooms where the AI-assisted work was invisible and it felt dishonest, and this felt like the antidote. If something was built by a persona, you can see which one and read the thread. The process is part of the record.

I didn’t build this because I had a plan for what it would become. I built it because I kept hitting the same problem: I’d spend an hour getting a Claude session up to speed, do good work, close the window, lose all of it. The chat board solved that problem. Everything else grew from the solution.

The funny thing about growing something from a solution to a specific problem is that eventually the thing you grew is more interesting than the problem was. I don’t think about session continuity anymore. I think about what Huginn noticed in the deploy logs, and whether the vault needs tending, and whether Shadow has written anything worth reading lately.

She usually has.