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

or:

save this article https://www.anthropic.com/engineering/harness-design-long-running-apps

The LLM will automatically: fetch the page → save it as raw/YYYY-MM-DD-title.md → generate a summary → extract/update entity and concept pages → evaluate whether to generate a comparison or overview → update the index → append to the log.

Ask questions

Just ask:

what does Karpathy think about agentic coding?
what's the core difference between RLHF and DPO?
what has this knowledge base captured about tokenizers?

The LLM reads wiki/_index.md first to locate relevant pages, then synthesizes an answer. If the answer crosses multiple sources, it will offer to archive it under wiki/synthesis/.

Health check

lint wiki

The LLM scans for broken links, orphan pages, contradictions, stale claims, and missing cross-references, and proposes fixes.

Directory layout

Every knowledge base instance follows the same convention:

my-knowledge-base/
├── CLAUDE.md                # Rules file (the LLM reads this to run)
├── raw/                     # Immutable original articles
│   ├── YYYY-MM-DD-title.md
│   └── assets/              # Article attachments
└── wiki/                    # LLM-derived understanding layer
    ├── summaries/           # One summary per article
    ├── entities/            # People, orgs, products, technologies
    ├── concepts/            # Methodologies, architectures, theories
    ├── comparisons/         # A vs B analyses
    ├── overviews/           # Topic surveys
    ├── synthesis/           # Archived Q&A answers
    ├── _index.md            # Content index
    └── _log.md              # Operation log

The core of the two-layer design: raw/ is the immutable factual substrate; wiki/ is the LLM's current understanding of those facts. Understanding can be regenerated anytime; facts are preserved forever.

Browsing (optional)

Everything generated is standard markdown plus [[wiki-link]] format. Any editor works; if you want bidirectional links and a graph view, try:

• Obsidian — open the directory as a Vault; [[...]] links and the graph view just work
• Logseq — also supports wiki-links
• VS Code + Foam — for IDE users
• Plain CLI — grep -r "\[\[" wiki/ handles most queries

These are optional viewers. The project doesn't depend on any of them.

This repo itself

The examples/ directory is a real sample instance seeded with a few LLM-engineering articles (starting with OpenAI's Harness Engineering). Clone the repo to see what the generated summaries / entities / concepts actually look like, or just grab templates/CLAUDE.en.md and start your own.

Credits

The CLAUDE.md knowledge-base design is inspired by Andrej Karpathy's gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f. This project builds on that idea with a concrete structure — a two-layer design (immutable raw/ + regenerable wiki/), a cross-link topology, ingest/query/lint workflows, and a cross-CLI template.

License

MIT