Breaking

Karpathy's 'LLM Wiki' Pattern Turns Obsidian and Claude Code Into an Agent-Maintained Second Brain

July 12, 2026 at 19:15 EDT

AI Agents
Software Dev & Coding
Open Source

Andrej Karpathy has popularized a lightweight blueprint for building a personal knowledge base in which a coding agent—rather than the user—compiles, links and maintains plain-text notes, reframing note-taking as a codebase managed by an AI. The approach, published as a GitHub Gist titled "llm-wiki", pairs the popular note app Obsidian with Anthropic's Claude Code, and has spread rapidly among developers experimenting with AI-assisted knowledge management.

llm-wiki · Agent-Managed Knowledge

Your notes become a codebase the AI maintains

Andrej Karpathy's lightweight blueprint pairs Obsidian with Claude Code: the agent — not you — compiles, links and audits plain-text notes, reframing note-taking as software an AI continuously refactors.

~100

articles in Karpathy's live vault

~400K

words managed as a knowledge codebase

41K+

stars on the companion obsidian-skills repo

core operations run the whole system

The three operations

Ingest

Compiles a raw source into structured, cross-linked Markdown pages.

→

Query

Answers from the generated wiki instead of re-searching sources.

→

Lint

Audits for contradictions, stale entries and orphaned pages.

Why it's different

Knowledge that compounds vs. snippets fetched and forgotten

Conventional RAG
re-fetches per query · no persistent structure

llm-wiki
persistent, interlinked wiki that grows more useful

All local Markdown, Git-versioned, RAG-free — no proprietary index or embedding store. raw/ holds immutable sources · wiki/ holds agent-generated pages · CLAUDE.md sets the schema.

What draws adopters

Knowledge compounds — links auto-maintained
Plain-Markdown, RAG-free simplicity
Reading notes structured by character & theme
Ingest call transcripts to surface patterns

Open questions

Single index page grows fragile at scale
LLM may hallucinate links or updates
Meaningful initial setup effort
Not a wholesale replacement for RAG

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.