I got tired of my Obsidian vault rotting, so I built an AI to maintain it. (Open-sourcing my LLM-Wiki)

At some point, I got incredibly frustrated with my own notes.

I had Obsidian vaults for five different ventures and client projects. Each one started strong and degraded within weeks. Notes became orphans. Cross-links rotted. The index was always out of date. I'd search for something I knew I'd written and find nothing—or three conflicting versions.

The standard productivity advice is "build a better system." I tried that. The real problem is that maintaining a knowledge base is administrative work that doesn't get done when you're busy building—which is all the time when you're running multiple things.

So I asked a different question: what if the LLM maintained it?

I recently saw Andrej Karpathy post a gist about the "LLM-Wiki" pattern—the idea of using a local AI to actively compile and manage a plain-text wiki. It was a brilliant baseline, so I took his method, integrated my own multi-venture workflows, and built a personal operating system out of it.

This isn't a chatbot that answers questions from outside your notes. It’s an agent that works inside the vault—ingesting sources, synthesising concept pages, maintaining cross-links, and running health checks. It operates via a plain-English schema that tells it exactly how the folder structure works and what to do with each workflow.

The result is LLM-Wiki. I've been using it locally as my daily driver, and it's the first setup that has actually stayed useful.

What it does

You give Claude Code a CLAUDE.md file at the vault root. It defines the folder structure, six core workflows, and a set of conventions (frontmatter, wikilinks, log format). When you open the terminal and run claude, it reads the schema and becomes the vault's maintainer.

Commands like "process the inbox", "promote this literature note", or "lint the vault" execute structured workflows. It creates literature notes from sources you feed it, synthesises them into permanent concept pages, and keeps an ops log of everything it touches.

Everything lives in plain Markdown you own. No proprietary format. No database. No sync subscription.

What I open-sourced today

The CLAUDE.md operating schema (the main thing — the pattern lives here)

9 note templates

6 worked examples (reading a book, watching a video, studying for a certification, ingesting a PDF, taking a course, reading tool docs)

A demo vault with a Deep Work concept page and a Cal Newport entity

Pre-configured .obsidian/ with Dataview + Templater registered

An optional Notion MCP sync skill for pushing actions out of the vault

It's MIT licensed, completely free, and there's no sign-up.

Honest limitations

It requires Claude Code (which runs on API credits, so it's not entirely free). It benefits from occasional human curation—it's a tool that amplifies your thinking, not a replacement for it. And it takes a session or two to get familiar with the workflow conventions.

Where this is going

I'm building other verticals on top of this open engine. The first is Founder OS—a calibrated, self-maintaining decision journal and operating system for founders (more details at founder-os.ozanyildirim.me). The engine stays open; the domain-specific skill packs and hosted layers are where the business goes.

I would love feedback from anyone who's tried to maintain a structured knowledge base over time. I'm curious to hear what broke for you and whether this pattern would solve it.

→ Vault Details: llm-wiki.ozanyildirim.me
→ GitHub: github.com/tyrozz/llm-wiki
→ The Karpathy gist that inspired the base architecture: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f