Karpathy-Inspired LLM Knowledge Base

English | 简体中文

One CLAUDE.md = a self-maintaining local knowledge base. No backend, no vector DB, no RAG framework.

curl -fsSL https://raw.githubusercontent.com/zhurudong/andrej-karpathy-llm-wiki/main/install.sh | bash -s my-kb

After that one line, open your LLM CLI inside my-kb/ and say ingest https://example.com/article — you now have a knowledge base the LLM organizes, indexes, and queries for you. Everything is plain markdown; open it with any editor.

Inspired by Andrej Karpathy's gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f.

Why

Most "personal knowledge base" solutions take one of two paths:

• Note-taking apps (Notion / Obsidian / Logseq) — great for storage and browsing, but tagging, linking, and organizing is all manual.
• RAG / vector search — great for Q&A over large corpora, but needs embedding services, a vector store, and an ingestion pipeline. Heavy, fragile, opaque.

This project takes a third path: let the LLM do the organizing, use markdown files as the substrate, use wiki-links as the graph, and use an LLM CLI as the runtime.

• Raw articles are immutable; LLM-generated summaries / entities / concepts can be recompiled anytime
• Everything is plain markdown — works with any editor, Git, grep
• The knowledge graph emerges naturally from [[wiki-link]] — no graph DB
• Switching LLM tools requires zero data migration — the rules live in CLAUDE.md

Typical use cases:

• Reading papers — drop an arXiv link; the LLM generates a summary and links it to existing concepts
• Following a field — ingest industry blogs regularly; overviews/ organically form topic surveys
• Archiving your own thinking — ask questions, let the LLM store synthesized answers in synthesis/, building your own opinion library
• Team collaboration — push to Git; teammates maintain the same knowledge base with their own LLM CLIs

Quick Start

One-line install (recommended)

One command bootstraps a fresh knowledge base — directory, CLAUDE.md, AGENTS.md symlink, and the empty raw/ + wiki/ skeleton:

curl -fsSL https://raw.githubusercontent.com/zhurudong/andrej-karpathy-llm-wiki/main/install.sh | bash -s my-kb

Pass a directory name as the first arg (defaults to my-knowledge-base). After it finishes, cd my-kb, launch your LLM CLI, and start talking.

Manual install

Prefer not to pipe a script into bash? Run these three commands instead:

mkdir my-knowledge-base && cd my-knowledge-base
curl -fsSL -o CLAUDE.md https://raw.githubusercontent.com/zhurudong/andrej-karpathy-llm-wiki/main/templates/CLAUDE.en.md
ln -s CLAUDE.md AGENTS.md

The single templates/CLAUDE.en.md is the entire "program" — it tells the LLM how to organize this knowledge base. The AGENTS.md symlink keeps the same file usable across CLIs:

CLI	Convention file
Claude Code	CLAUDE.md
Codex CLI	AGENTS.md (symlink to CLAUDE.md)
OpenCode	AGENTS.md
Other agent CLIs that read a project rules file	see their docs

Start ingesting

Inside your LLM CLI, just use natural language:

ingest https://www.anthropic.com/engineering/harness-design-long-running-apps