An LLM Wiki is a pattern for persistent, compounding knowledge bases maintained entirely by LLMs. The human discovers sources and asks strategic questions; the LLM handles all bookkeeping: summaries, cross-references, contradiction resolution, and index maintenance. The pattern was proposed by Andrej Karpathy in April 2026.
The Core Insight
LLMs excel at mechanical bookkeeping that humans reliably abandon. Most knowledge bases die not from lack of input but from lack of maintenance. Cross-references go stale, indexes become incomplete, contradictions accumulate unnoticed, and new material is never integrated with existing content. LLMs eliminate this failure mode by performing maintenance as a side effect of every operation.
Three-Layer Architecture
Layer 1: Raw Sources
User-curated, immutable documents: articles, papers, repos, data. The LLM reads this layer but never modifies it. These are the primary sources that ground the wiki in reality.
Layer 2: The Wiki
LLM-generated and LLM-maintained Markdown files: entity pages, concept pages, summaries, and interconnected analyses. The LLM owns this layer entirely. It creates, updates, cross-links, and maintains articles as new sources are ingested and queries are answered.
Layer 3: The Schema
A configuration document (e.g., CLAUDE.md, AGENTS.md) defining structure, conventions, and workflows. This is Context-as-Code applied to knowledge base maintenance. The schema tells the LLM how to organize, name, tag, and cross-reference articles.
Core Operations
Ingest — When a new source is added, the LLM extracts key takeaways, writes or updates wiki articles, creates cross-references to existing articles, and appends to the operation log.
Query — The LLM searches relevant articles, synthesizes an answer with citations, and optionally promotes valuable query outputs into new wiki pages.
Lint — A health-check operation that scans for contradictions, stale claims, orphan articles, missing cross-references, and index inconsistencies. Lint repairs what it finds.
Structural Files
Every LLM Wiki has two mandatory structural files:
- Index — A content-organized catalog with one-line summaries for each article. Updated on every ingest, explore, or lint operation.
- Log — A chronological, append-only operation record with parseable prefixes. Provides full audit trail.
Why File-Based Beats RAG
At scale (~100 articles / ~400K words), a well-maintained wiki with auto-maintained indexes outperforms RAG for question-answering. The LLM navigates file structure directly — following cross-references and reading targeted articles — rather than relying on embedding-based retrieval that can miss nuance and context.
The Farzapedia variant (by Farza) demonstrated this with 2,500 diary entries producing 400 articles with backlinks. His key distinction: the wiki is built for the agent, not the human. File-system structure with backlinks is more easily crawlable than vector similarity search.
LLM Wiki and Compounding Knowledge
LLM Wikis are a concrete implementation of Compounding Knowledge. Every ingested source triggers updates across multiple existing articles. Every query can generate new articles or improve existing ones. The wiki becomes a compounding artifact where the act of using it makes it more valuable.
This is also how Agentic Knowledge Management manifests in practice: the AI agent does not just answer questions — it builds and maintains the knowledge infrastructure.
Key Points
- LLMs handle the bookkeeping humans reliably abandon: cross-references, indexes, contradiction detection
- Three layers: raw sources (immutable), the wiki (LLM-maintained), the schema (conventions)
- Core operations: ingest, query, lint
- File-based navigation outperforms RAG at scale for structured knowledge
- The wiki compounds in value with every operation
Open Questions
- How to handle conflicting information across sources?
- What is the optimal article granularity for LLM consumption?
- Can multiple humans curate a shared LLM Wiki effectively?
References
- Andrej Karpathy, "LLM Wiki" (April 2026)
- Farza, Farzapedia implementation
- Vault: LLM Wiki, Compounding Knowledge