llm-wiki.md

Andrej Karpathy’s proposal for a persistent, compounding personal knowledge base maintained by LLM agents.

Core Thesis

Traditional RAG (Retrieval-Augmented Generation) is “lossy” because the LLM must re-discover and re-synthesize knowledge from scratch on every query. The llm-wiki pattern advocates for incremental compilation: when a new source is added, the LLM integrates it into a structured, interlinked wiki.

Architecture

  1. Raw Sources: Immutable truth.
  2. The Wiki: Compiled, LLM-maintained markdown files.
  3. The Schema: Operational instructions (AGENTS.md).

Operations

  • Ingest: Extracting insights from raw sources into the wiki.
  • Query: Using the compiled wiki to answer questions with context.
  • Lint: Health-checking for contradictions and broken links.

Connection to this Project

This Digital Brain is a live implementation of the llm-wiki pattern, using Quartz for the UI and an autonomous agent as the maintaining agent.