my-llm-wiki

Drop any files into a folder. Get a living, queryable knowledge graph.

Documentation · Issues

---

In April 2026, Andrej Karpathy shared a concept he called LLM Wiki — a personal knowledge system with three layers: raw files (never modified), a compiled wiki with cross-references, and a schema that tells the LLM how to maintain it. The key insight: compile once, query forever, and let the wiki grow with every session as a "persistent, compounding artifact" rather than re-deriving knowledge on every query.

my-llm-wiki implements all three layers. See How It's Built for the full narrative on how Karpathy's vision is realized.

pip install my-llm-wiki
cd your-project && llm-wiki .

The Living Wiki

One command builds the graph. The living wiki cycle keeps it growing over time — each session adds knowledge, insights get filed back, the graph compounds.

┌──────────────────────────────────────────────────┐
│                                                  │
│   ┌──────────┐    ┌──────────┐    ┌──────────┐  │
│   │ Monitor  │───▶│ Rebuild  │───▶│  Lint    │  │
│   │ (watch)  │    │ (cached) │    │ (health) │  │
│   └──────────┘    └──────────┘    └──────────┘  │
│        ▲                               │         │
│        │                               ▼         │
│   ┌──────────┐                  ┌──────────┐    │
│   │  Report  │◀─────────────────│Write-back│    │
│   │ (stats)  │                  │(insights)│    │
│   └──────────┘                  └──────────┘    │
│                                                  │
└──────────────────────────────────────────────────┘

Two passes extract knowledge from any file type:

Pass	What	Cost
Structural	AST (18 languages): classes, functions, typed extends/implements edges, function signatures, doc comments (Javadoc/JSDoc/GoDoc), call graph, headings, cross-ref	Free
Semantic	Claude Code agents read DOCX, scanned PDFs, images with vision	Claude tokens

Output goes to wiki-out/:

wiki-out/
  graph.html       ← interactive graph (vis.js)
  graph.json       ← persistent graph data
  WIKI_REPORT.md   ← god nodes, surprising connections
  wiki/            ← Wikipedia-style articles
  vault/           ← Obsidian-compatible vault ([[wikilinks]] + YAML frontmatter)
  cache/           ← SHA256 cache (skip unchanged files)

Obsidian integration

wiki-out/vault/ is a drop-in Obsidian vault. Open it directly, or symlink into an existing vault:

llm-wiki .
# Obsidian → Open folder as vault → wiki-out/vault/

You get: graph view (force-directed), backlinks, tag pane, full-text search, and Properties view (Obsidian 1.4+ reads the YAML frontmatter on each node). Community colors are pre-configured via .vault/graph.json. Use llm-wiki query from CLI for typed-edge details (Obsidian wikilinks are untyped, so extends/implements/calls collapse to generic links in the Obsidian graph view).

CLI

llm-wiki .                          # build graph
llm-wiki query gods                 # most connected nodes
llm-wiki query search <term>        # keyword search
llm-wiki query path <A> <B>         # shortest path
llm-wiki lint                       # graph health check
llm-wiki watch .                    # auto-rebuild on changes
llm-wiki add <url>                  # ingest URL
llm-wiki note "<insight>"           # write-back from LLM session

Claude Code Skill

mkdir -p ~/.claude/skills/my-llm-wiki
cp "$(python -c 'import my_llm_wiki; print(my_llm_wiki.__path__[0])')/SKILL.md" ~/.claude/skills/my-llm-wiki/

Then /wiki . in Claude Code — structural extraction + agent-mode semantic extraction for DOCX, scanned PDFs, images.

Docs

phuc-nt.github.io/my-llm-wiki

License

MIT