AI Second Brain

July 2026

A personal knowledge base where I bring the material and an AI agent does the writing. I feed it articles, videos and my own raw thoughts; the agent turns them into wiki pages, links everything together and keeps it all up to date. I browse the result in Obsidian like a map of what I know.

The idea

Most note-taking systems die the same way: saving things is easy, but keeping them organized, linked and current is tedious, so the notes go stale and nobody reads them again. The work that kills a second brain is exactly the work an AI agent is good at.

So I split the roles. The project follows the LLM Wiki pattern shared by Andrej Karpathy: instead of searching raw documents every time (as classic RAG does), an agent builds and maintains a lasting wiki of Markdown pages, so knowledge adds up instead of being rediscovered.

What I do
  • Choose what is worth keeping
  • Write my thoughts freely, however messy
  • Ask questions
  • Decide the judgment calls
What the agent does
  • Read and summarize every source
  • Write and update all the wiki pages
  • Link related pages in both directions
  • Keep the index and the history log current

Three layers

The vault is a folder of Markdown files with three layers, each with a clear owner:

Raw · mine Sources Articles, transcripts, images and journal entries, exactly as they arrived. Never edited afterwards: they are the source of truth.
→read by the agent
Wiki · the agent's Knowledge Every page written and maintained by the agent, cross-linked with Obsidian wikilinks.
areas/entities/concepts/sources/notes/index.mdlog.md
Schema · the rules AGENTS.md One file describes how the agent must behave: page templates, frontmatter, linking rules and every workflow. It turns a general-purpose AI into a disciplined librarian.

A map of connected ideas

Because the agent links every page to the pages it relates to, the vault grows into a network rather than a pile of notes. This is its real graph in Obsidian: each dot is a page, each line a link the agent wrote, and the biggest dots are the hubs that many ideas lead back to.

Obsidian graph of the knowledge base: about a hundred linked pages, with a few large hubs

The vault's graph in September 2026, about three months in. Page names are hidden to keep its content private.

How I use it

Ingestadd something new
  1. I clip an article with the Web Clipper extension, or write a thought
  2. We discuss the key takeaways first
  3. The agent writes a summary page
  4. It updates every related page
  5. Index and log updated, then a commit
Queryask a question
  1. The agent starts from the index
  2. It reads the relevant pages
  3. It answers with citations
  4. A good answer is filed back as a new note
Linta regular health check
  1. Contradictions never reconciled
  2. Claims made stale by newer sources
  3. Orphan pages and missing links
  4. Fixes the obvious, asks me about the rest

Philosophy

Sources never changeWhat came in stays as it came in. Corrections and new understanding live in the wiki, never in the originals.
Changing my mind stays visibleWhen something new contradicts an old page, the agent notes the contradiction instead of silently overwriting. How my thinking evolves is the most valuable part.
Opinions are labelled as opinionsAnything that comes from my own thoughts is marked as my view, with a date, so it never reads like an established fact.
Stay in the loopThe agent discusses each source with me before filing it, rather than quietly rewriting a dozen pages on its own.
Synthesize, don't transcribeShort pages that are correct and current beat long pages that are stale.
Answers are worth keepingA good analysis should not disappear into chat history: it becomes a page and feeds future questions.
No lock-inPlain Markdown and one vendor-neutral rule file: any AI agent can take over tomorrow, and the notes outlive every tool.
Private by designThe vault lives in a private repository and its browsing site only runs on my machine. This page describes the system, never its content.

Tech stack

ObsidianReading and browsing, with the graph view and the Extended Graph plugin to see how ideas connect.
Obsidian Web ClipperA Chromium browser extension: one click saves the article I'm reading into the sources folder as clean Markdown, ready for the agent to process.
AI coding agentsClaude Code, Gemini CLI or Codex. Each reads the same AGENTS.md, through a tiny stub file when it expects another name.
Markdown + YAML frontmatterEvery page has a type, an area, tags, dates and the sources behind its claims.
Git + GitHubHistory of every change, and sync between my machines: pull before working, push after.
QuartzA local website built from the vault, with search, graph and backlinks in the browser.

Inspiration

LLM Wiki, by Andrej Karpathy
The idea file behind this project: let an LLM build and maintain a wiki instead of re-reading raw documents.
gist.github.com