Karpathy LLM Wiki Pattern

Source: Andrej Karpathy’s Gist
Pattern name: LLM Wiki
Published: ~2022
Status: active

Core Claim

Store knowledge as natural language documents. Let LLMs handle retrieval and synthesis. The machine learns to find relevant content by reading everything — no manual indexing, no rigid schema.

Key Properties

  • Compile once, retrieve always — knowledge is compiled into a searchable corpus
  • No manual tagging — LLMs navigate by understanding content, not metadata
  • Schema lives in the prompt — not in the data
  • Natural language storage — documents as the unit of knowledge

How It Maps to This Vault

The OpenClaw wiki uses a similar principle:

  • Sources = natural language documents (compiled knowledge)
  • Concepts = claims extracted from sources (structured understanding)
  • The LLM reads both to answer questions

Contrast With Traditional Knowledge Management

Feature Traditional KMS LLM Wiki Pattern
Organization Manual hierarchy Emergent from content
Retrieval Tags + folders Semantic search
Schema Rigid frontmatter Prompt-defined
Updates Labor-intensive Append-only corpus

See Also

  • No related pages yet.