The Day My Vault Started Talking Back
I have been keeping a digital notebook for years. Obsidian vaults, markdown files, folders within folders — the usual second brain setup. But there was a problem: my vault was a library, not a brain. It stored information, but it never talked back. That changed when I discovered the Karpathy/Wolfe system for AI-powered knowledge management.
Andrej Karpathy first described the concept of an LLM-powered wiki — a knowledge base where an AI agent reads, summarizes, and links your content so you can query it naturally. Matt Wolfe expanded it into a three-pillar system: a wiki for processed knowledge, a journal for daily experiences, and a CRM for people you interact with. The key insight? The AI doesn’t just store your notes — it grounds its answers in your own vault, not just in its training data.
The Three Pillars
I restructured my vault around three core directories. The wiki folder holds processed knowledge — articles I clipped from the web, YouTube transcripts, technical documentation. My AI agent (Hermes) reads each new entry, extracts concepts, links them to existing pages, and builds a growing web of interconnected ideas. The journal folder captures daily logs — not just what happened, but what I learned, what broke, what I fixed. And the CRM tracks people: who I talked to, what we discussed, what tools or ideas they shared.
The magic happens when these three pillars connect. A journal entry about a Hermes upgrade links to a wiki page about the Hermes Agent. That wiki page links to a CRM entry for the developer who built a related tool. Suddenly, my vault is not a collection of files — it is a knowledge graph that my AI can traverse.
What I Learned
The biggest lesson: a second brain that never talks back is just a fancy filing system. The AI layer is what transforms a vault into something alive. When I ask Hermes “What did I learn about Obsidian CLI tools last month?” it doesn’t guess — it searches my wiki, finds the relevant pages, and answers from my own knowledge. That is radically different from asking ChatGPT, which answers from the internet’s knowledge.
I also learned that structure matters more than volume. A vault with 300 well-linked pages is more useful than a vault with 3000 orphaned notes. The Karpathy system forces you to process, not just collect. Every new piece of content goes through a pipeline: clip → summarize → extract concepts → link to existing pages. It takes discipline, but the compound effect is real.
Try It Yourself
You do not need a complex setup. Start with one folder — wiki, journal, or CRM — and one AI tool that can read your vault. Obsidian Web Clipper makes it easy to capture content. An AI agent like Hermes or Claude Code can process it. The key is consistency: process one piece of content per day, link it to something existing, and watch your knowledge graph grow.
Have you tried building an AI-powered second brain? What tools are you using? Drop a comment below — I would love to hear what works for you.
Wat vond je van dit bericht?
