MindBase — Karpathy's LLM Wiki, as a product

An open-source implementation of Andrej Karpathy's LLM Wiki idea: an AI that builds and maintains a wiki from your sources. Not RAG-in-a-vector-DB. A real markdown wiki on your disk, that an LLM gardens for you between conversations.

npx mindbase-app

One command: starts the local server, opens the web app, and walks you through picking a free local model that fits your RAM. No API key, nothing leaves your machine. (Node 20+)

MindBase implements Andrej Karpathy's LLM-Wiki pattern: you feed it sources (papers, articles, thoughts); the LLM reads, cross-references, flags contradictions, and writes structured wiki pages. Later, when you ask a question, the wiki already has the synthesized answer — no vector-search re-derivation at query time.

Status: Early access, actively developed. What's new: CHANGELOG · Releases

Why MindBase

You read a lot. Papers, articles, tweets, docs. You want to remember them, connect them, form opinions from them. Today you have two bad options:

• Notion / Obsidian / Roam: Passive containers. You do all the organizing. AI features are bolted-on generation, not maintenance.
• NotebookLM / Perplexity Pages / ChatGPT search: RAG-based. Nothing accumulates. Every question re-derives the answer from raw sources.

MindBase is the third option: the LLM actively maintains a persistent, structured wiki as you feed it sources. Knowledge compounds. Your context.md gets sharper every time you contribute. The AI remembers you across sessions because your beliefs are written down in markdown files — not stored in a chat history that gets summarized away.

Think of it as a personal Wikipedia that an AI intern writes for you, kept up to date, cross-referenced, and honest about what it doesn't know.

How it works (30 seconds)

Three physical layers on disk (Karpathy's model):

Three operations run the loop: ingest (AI reads a source, discusses takeaways, you approve the wiki updates), build (regenerate context.md from everything unbuilt), lint (the AI audits its own wiki for contradictions, stale claims, and orphans).

What makes it different

• The AI asks before writing. Every ingest shows takeaways + a checkbox plan; only what you approve gets written. No black-box edits to your knowledge.
• You can watch the wiki absorb your notes. Every note carries a status chip — ✨ Add to wiki until digested, ✓ In wiki after.
• It audits its own knowledge. One command re-reads the whole wiki and reports contradictions with the exact conflicting sentences quoted. Notion and NotebookLM structurally cannot do this.
• Free and local by default. Hardware-detect wizard installs the best Ollama model for your RAM. Cloud keys optional.
• Plain markdown on disk. Grep it, git it, open it in Obsidian, leave anytime.

Install

Browser (fastest): npx mindbase-app — shown above. Everything runs locally at localhost:4321.

Claude Code (flagship): the full Karpathy 8-step ingest with sub-agents, slash commands, and per-agent tool boundaries:

/plugin marketplace add frankchu91/mindbase-llm-wiki
/plugin install mb@mindbase

Restart when prompted, then type / — you should see /mb:contribute, /mb:build, /mb:ask, /mb:lint and 8 more. You get 5 sub-agents with strict tool allowlists (the builder has no file-write tool at all — only an atomic-write MCP call), plus a SessionStart hook that auto-injects your project context.

{ "mcpServers": { "mindbase": { "command": "npx", "args": ["-y", "mindbase-mcp"] } } }

{ "mcpServers": { "mindbase": { "command": "npx", "args": ["-y", "mindbase-mcp"] } } }

{ "mcpServers": { "mindbase": { "command": "npx", "args": ["-y", "mindbase-mcp"] } } }

{ "experimental": { "modelContextProtocolServers": [ { "transport": { "type": "stdio", "command": "npx", "args": ["-y", "mindbase-mcp"] } } ] } }

Next: create your first project and learn the four daily workflows in the Guide →

The web UI

Since 0.3 the browser app stands on its own — write notes in a full WYSIWYG editor (Cmd+N), quick-capture from anywhere (Cmd+I), and run the AI operations with approval cards: /contribute, /build, /lint, /research. Live demos on the website.

Free local models: the setup wizard detects your hardware and installs what fits — llama3.2:3b (8GB), qwen3:14b (24GB+), or Meta's Muse Glimmer 30B (32GB+ Apple Silicon, Ollama ≥ 0.32.7). Measured guidance: qwen3:14b for interactive work (~30s), Glimmer for background lint/build — slower, but its findings quote the exact conflicting sentences. The model switcher on the chat composer flips between them in two clicks.

Where your data lives

Everything is plain markdown under ~/mindbase-data/ (override: MINDBASE_DATA_DIR):

~/mindbase-data/projects/my-research/
├── README.md                 # Ops manual — you edit, LLM reads
├── context.md                # Synthesized truth — LLM writes, you approve
├── index.yaml                # Auto-generated catalog
├── sources/
│   ├── contributors/<you>/   # Your dated entries + notes (append-only)
│   ├── research/             # LLM-authored wiki pages
│   └── raw/                  # PDFs, HTML captures
├── logs/                     # Chronological operation log
├── artifacts/                # Briefs, exports, lint findings
└── state/builder/snapshots/  # context.md snapshots for rollback

No proprietary database — what you see on disk is what MindBase knows. git init it, back it up with anything, delete a project with rm -rf.

Architecture at a glance