Karpathy LLM Wiki Source

karpathy-llm-wiki-source (commonly referred to as the LLM Wiki pattern) is a conceptual methodology for personal knowledge management (PKM) introduced by Andrej Karpathy in April 2026. Rather than relying on traditional Retrieval-Augmented Generation (RAG), which retrieves and synthesizes information dynamically for each query, this pattern utilizes an AI Agent to actively and incrementally “compile” raw information into a structured, persistent, and interlinked set of Markdown files.

Core Philosophy: “Compile, Don’t Retrieve”

In standard RAG systems, the LLM reads chunks of raw documents to answer questions but does not build long-term, evolving associations, effectively starting from scratch every time. The LLM Wiki pattern introduces a shift by treating knowledge as a compiled, structured codebase:

• Accumulative Knowledge: New source materials (PDFs, notes, articles) are read by the AI agent, which updates existing entity files, creates cross-references, and integrates new insights.
• Conflict Resolution: The agent actively flags contradictions or discrepancies between new and existing sources.

Key Features

• Active Compilation & Maintenance: The knowledge base is updated continuously or on-demand by an agent following a defined schema.
• Self-Healing/Linting: The agent runs checks for orphaned pages, stale claims, or broken links to maintain organization.
• Interlinked Structure: Uses bidirectional [[wiki-links]] to build a semantic map of concepts.
• Portability & Human-Readability: Because the output consists of plain Markdown files, it remains accessible, future-proof, and can be viewed using frontend tools like Obsidian.

Typical Architecture

The pattern is implemented using three main directories/configuration layers:

1. Raw Sources (raw/): A directory containing immutable input documents (transcripts, papers, articles).
2. The Wiki (wiki/): A collection of structured, cross-linked Markdown files representing the compiled knowledge base.
3. The Schema (CLAUDE.md): A configuration file detailing instructions for the AI agent (e.g., Claude Code) on how to ingest files, link entities, and format updates.

Related Topics

• Retrieval-Augmented Generation (RAG)
• Agentic Workflows
• Personal Knowledge Management (PKM)
• Semantic Knowledge Graphs