Lately I've been hearing a lot about AI that works from a deep store of knowledge rather than improvising each answer. The idea is simple: connect the things you know, and let the AI draw on them. Isn't that what we were promised in the first place? Seems to be the same regardless on hierarchy.

I recognized it straight away, because I've been building a small version at home in my homelab, using Hermes hosted on a Proxmox system, and it turned out to be much easier than I expected. It also shows that everyone, whatever their size, runs into the same limit: you can't simply pour your data into a model and expect it to know it. That's why the idea of a "second brain" keeps coming up. Can't just tell your minions to figure things out, well they will as we already proved time and time again, a leader should know this and not find out on the fly because of who is underneath them.

Sapiens Brain was a oversight! Not sure if we were waiting for Anthropic to give us the perfect modal with the knowledge we accumulated?

The company I work for has its own take on this with Sapiens Brain, so it was funny to find I'd already been building a small personal version. Mine is simply a collection of notes that link to each other, plus an AI that can read, search and add to them. Everything it says is grounded in notes I've written or approved.

What mine looks like

- The notes live in an Obsidian vault, which is just a folder of text files. They sync to my own server through NextCloud, so they're on my laptop and phone.

- Currently three AI agents share that vault, each with a different job: one for the home, one that looks after the notes, and one for my website.

- A memory layer helps the note-keeper agent remember things between sessions, so I don't repeat myself.

- A 3D graph shows the notes, the agents, and which agents are reading or writing which notes. Notes are by topic, so I can see at a glance where the knowledge is thick and where it's thin.

Every piece is ordinary software. Nothing here needed a research team.

What I learned

The hard part isn't the technology. Choosing what goes into the vault, and what stays out, took more thought than any of the setup. I decided early that work material stays out of my home system entirely. Funny watching how everyone above me are doin this.

Structure matters more than volume. Fifteen well-linked notes beat a thousand loose ones. The links are what let an AI find the right answer rather than a plausible one. Also matters that no matter what amazing modal is provided for us, its not going to know our system! That's why we always relied on what we have stored on Jira and Confluence, that is our 'source' of information.

Consistency comes from grounding. When an agent answers from filed notes instead of guessing, the answers stop drifting, and I can open the note and check the source. It also taught me some humility. Nobody has the full picture of AI, and you can't walk in assuming you know everything and tell others what to do. It's moving fast and I'm still learning which is what we all are doing, thrown ourselves into the deep end to figure it all out.

It exposes your gaps. The graph shows you quickly which topics have notes and which don't.