Astro-Han/karpathy-llm-wiki

What it solves

It addresses the limitation of traditional RAG (Retrieval-Augmented Generation), where an LLM must re-search and re-synthesize raw document chunks every time a question is asked. Instead of relying on ephemeral retrieval, this project implements a system where an LLM maintains a durable, structured wiki of synthesized knowledge that compounds and improves over time.

How it works

The project is implemented as an "Agent Skill" that allows coding agents (like Claude Code or Cursor) to manage a knowledge base using three primary operations:

• Ingest: Takes a source (URL, file, or text), stores it in a raw/ directory, and then creates or updates synthesized markdown pages in a wiki/ directory.
• Query: Searches the existing wiki pages to provide grounded answers with direct citations to the markdown files.
• Lint: Scans the wiki for broken links, missing index entries, and stale cross-references to ensure health and consistency.

Who it’s for

It is designed for users of AI coding agents and LLM tools who want to build a compounding knowledge base where the AI handles the curation, summarization, and cross-linking of information from various sources.

Highlights

• Durable Knowledge: Shifts synthesis from query-time (RAG) to ingest-time (Wiki).
• Agent-Maintained: The LLM handles the updating of summaries and cross-references as new material arrives.
• Tool Agnostic: Compatible with the Agent Skills standard, working across tools like Claude Code, Cursor, and Codex CLI.
• Simple Architecture: Uses a basic file-system structure (raw/ for immutable sources and wiki/ for curated pages) rather than complex vector databases for small-to-medium knowledge bases.