A second brain that actually runs unattended
Notes on building an agentic system that holds project state across dozens of threads — self-hosted, and used every day.
The hardest problem on a large nuclear project is rarely the engineering.
It is keeping track of who did what, when, what is still outstanding, and how a hundred moving pieces line up against the schedule. I lived that problem for three decades at Darlington, most recently coordinating scheduling across seventeen concurrent projects during refurbishment.
Why a chatbot doesn’t solve it
A chatbot answers a question and forgets you asked. That is fine for drafting an email and useless for holding project state.
What is needed is something that persists — that knows what was decided in March, what is still blocked, who owns it, and what changed since you last looked. The difference is not model quality. It is architecture.
What I actually run
The system lives on my own hardware:
- A three-node Proxmox cluster
- A K3s cluster for the container workloads
- TrueNAS underneath it for storage
- Everything reachable only over the LAN or Tailscale — no public exposure
It tracks tasks, project state, documentation and follow-through across dozens of active threads. It is not a demo. I use it every day, and most of what I have learned came from the parts that broke.
More on those parts in the posts that follow.