concept
LLM Wiki
LLM Wiki
The foundational pattern behind the Kulify Second Brain. Proposed by Andrej Karpathy as an alternative to RAG.
Core Idea
Instead of re-deriving knowledge from raw documents on every query (RAG), have the LLM incrementally build and maintain a persistent wiki. Knowledge is compiled once, kept current, and compounds over time.
Three Layers
- Raw Sources — Immutable documents (articles, papers, data) that the LLM reads but never modifies
- The Wiki — LLM-generated markdown files (summaries, entity pages, concept pages, comparisons). The LLM owns this layer entirely
- The Schema — A configuration document (CLAUDE.md) that defines structure, conventions, and workflows. Co-evolved by human and LLM
Three Operations
- Ingest: Drop a new source → LLM reads it, writes summary, updates 10-15 related wiki pages
- Query: Ask questions → LLM reads relevant pages, synthesizes answers. Good answers get filed back
- Lint: Periodically health-check for contradictions, stale claims, orphan pages, missing cross-references
Why It Works
Humans abandon wikis because maintenance burden grows faster than value. LLMs eliminate that cost. The human curates sources, directs analysis, and asks good questions. The LLM handles all bookkeeping.
Karpathy likens this to Vannevar Bush’s 1945 Memex vision — a personal knowledge store with associative trails — where the LLM finally solves the maintenance problem.
Navigation Aids
index.md— Content catalog for finding pageslog.md— Chronological append-only record of operations
Recommended Tools
- Obsidian as viewer/IDE
- CLI-based LLM agent (Claude Code, Codex)
- Obsidian Web Clipper for converting articles to markdown
- Dataview for querying page frontmatter
- Git for version history
How We Use It
The kulify Second Brain implements this pattern directly. See CLAUDE.md for the schema.
Related
- Autoresearch — Karpathy’s other AI automation project
- Second Brain — broader concept