LLM Wiki Pattern
Source: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f Author: Andrej Karpathy Date: 2025
Summary
Karpathy proposes a three-layer architecture for LLM-maintained personal knowledge bases: Raw Sources (immutable), The Wiki (LLM-maintained), and The Schema (instructions). Core insight: instead of RAG (retrieve-at-query-time), have LLMs incrementally build a persistent wiki that compounds over time. This is the pattern this wiki is based on.
Key Claims
- Traditional RAG is stateless; the wiki is a persistent, compounding artifact
- Cross-references are pre-built; contradictions are pre-flagged — query-time is fast
- Three workflows: Ingest (process new sources), Query (search wiki, file results), Lint (audit contradictions)
- Humans curate sources and ask questions; LLMs handle all bookkeeping
- Applies to personal research, business teams, competitive analysis, hobby deep-dives
- The schema is configuration — different schemas for different use cases
Why It Differs from RAG
- RAG rebuilds context at query time → expensive, no compounding
- Wiki approach pre-processes and cross-references → answers are already partly synthesized
- Wiki surfaces contradictions proactively vs. RAG giving inconsistent answers from conflicting sources
Entities
Concepts
- Agent Memory — wiki as externalized LLM memory
- Context Engineering — the schema as context configuration