# LLM Wiki
An LLM Wiki is a pattern for building persistent, compounding knowledge bases maintained entirely by [[Large Language Models (LLMs)]]. The concept was proposed by [[Andrej Karpathy]] in April 2026. Instead of humans manually curating notes, the LLM handles all the tedious bookkeeping: writing summaries, maintaining cross-references, resolving contradictions, and keeping an index up to date. The human's role shifts to source discovery and strategic questioning.
## Three-Layer Architecture
The system is organized in three layers:
- **Raw Sources**: an immutable collection of documents (articles, papers, repos, datasets, images). The LLM reads but never modifies these
- **The Wiki**: LLM-generated [[Markdown]] files organized as entity pages, concept pages, summaries, and interconnected analyses. The LLM owns maintenance entirely
- **The Schema**: a configuration document (e.g., CLAUDE.md, AGENTS.md) that defines wiki structure, conventions, and workflows, guiding disciplined LLM operation
## Core Operations
- **Ingest**: process new sources by extracting takeaways, writing summaries, updating cross-references across existing pages, and appending to a chronological log
- **Query**: search relevant wiki pages and synthesize answers with citations. Valuable outputs become new wiki pages, compounding knowledge over time
- **Lint**: health-check the wiki for contradictions, stale claims, orphan pages, and missing cross-references
## Special Files
The wiki relies on two key structural files:
- **index.md**: a content-organized catalog listing each page with one-line summaries and metadata, organized by category
- **log.md**: a chronological append-only record of all operations, with parseable prefixes (e.g., `## [2026-04-02] ingest | Article Title`)
## Tooling
Karpathy uses [[Obsidian]] as the "IDE frontend" for viewing raw data, the compiled wiki, and derived visualizations. The [[Obsidian Web Clipper]] extension converts web articles into markdown files for ingestion. Additional vibe-coded tools include a naive search engine (usable via web UI or CLI by the LLM).
## Key Insight
LLMs excel at the mechanical bookkeeping that humans reliably abandon: updating references, maintaining consistency across hundreds of pages, noting contradictions, and imputing missing data. This frees humans to focus on curation, analysis, and asking the right questions. The wiki becomes a compounding artifact where cross-references and synthesis accumulate without fatigue.
## Critique: Compilation vs Curation
Steven Thompson pushes back on the LLM Wiki pattern from a [[Personal Knowledge Management (PKM)]] standpoint: it risks confusing **compilation** with **curation**. A compiler can maintain structure; only a human can maintain significance — sensing the *weight* of an idea, recognizing when understanding shifts, identifying which fragments are formative versus merely interesting.
The reframe: a PKM system is a **biography of understanding**, not a database. The aim isn't a perfectly coherent repository of everything one has read; it's a record of how one's thinking changes over time. Automate the maintenance and one risks automating away the meaning-making itself. As Thompson puts it: *"I'm not building a repository of everything I know; I'm building a record of how my understanding changes."*
This raises the **delegation question**: how much of the bookkeeping can be handed to the LLM before the wiki stops being *mine* and becomes a tidy, generic compilation indistinguishable from anyone else's? The friction of manual curation — choosing, weighing, connecting — is part of the work, not an obstacle to it.
A useful heuristic when ingesting into an LLM Wiki: distinguish *information* from *intellectual lineage*. Information belongs in a compilation; only intellectual lineage belongs in a biography of understanding. In practice, this argues for hybrid use — let the LLM compile the breadth (entity pages, summaries, cross-references) while reserving the [[Zettelkasten method]] core (permanent notes, folgezettel chains, weighted connections) for human curation.
## Relation to RAG
Karpathy notes that at the scale of ~100 articles / ~400K words, a well-maintained wiki with auto-maintained index files outperforms [[Retrieval-Augmented Generation (RAG)]] for Q&A. The LLM can navigate the file structure directly rather than relying on embedding-based retrieval.
## Wiki vs search (Anthropic, November 2026)
[[Anthropic]] tested the pattern at a much larger scale in its research post "The components of agent memory". Over 128 sessions, Opus 5 agents built a wiki from about 350,000 internal Stainless files (~420M tokens): they wrote and linked pages, merged duplicates and checked claims against the sources, with a shared log. On 25 eval questions (share of graded criteria met):
| Setup | Result |
|---|---|
| Wiki only | 13% → 48% as the wiki grew |
| Wiki + search | 54% → 61% |
| Search only | 53%, flat |
Wiki plus search also saved about 20,000 tokens per question. The wiki grew from 103 pages and 766 links after 2 sessions to 222 pages and 12,893 links after 128. On 24 Harvey LAB merger questions that need several documents, the wiki scored 87% against 73% for search. When the criteria sat in a single document, the wiki added nothing.
This nuances Karpathy's claim above. At that scale, the wiki did not replace search; it helped on top of it, and only for questions that span several documents.
Two failure modes showed up as well. Bloat: one wiki weighed 203 KB for a 115 KB corpus (1.77 times the source), with orphan pages. Dangling links: sessions linked to pages that didn't exist, and a second pass repointed nine links and created three pages. That's the Lint operation, which Anthropic runs as a "dreaming" pass (see [[Agent Memory Hierarchy]]). Parallel writers are a risk too: agents writing to one shared page kept only 41 of 57 facts, until the wiki instructions switched to one page per contract.
## Farzapedia
Farza (FarzaTV) independently built a similar system called [[Farzapedia]]: a personal Wikipedia generated from 2,500 diary entries, Apple Notes, and iMessage conversations, producing 400 detailed articles with backlinks. His key insight is that the wiki is built for the agent, not the human. The file-system structure with backlinks is easily crawlable by any agent, making it a more effective [[Knowledge Graph (KG)]] than RAG-based approaches.
## Future Direction
Karpathy envisions synthetic data generation and fine-tuning to embed wiki knowledge directly into model weights rather than relying solely on context windows. He also sees room for a dedicated product rather than a collection of scripts.
## Quotes
- [[Tools to enhance understanding]] — [[Andrej Karpathy]]
## References
- https://x.com/karpathy/status/2039805659525644595
- https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
- https://x.com/omarsar0/status/2040099881008652634
- https://x.com/FarzaTV/status/2040563939797504467
- https://academy.dair.ai/blog/llm-knowledge-bases-karpathy
- https://medium.com/a-voice-in-the-conversation/the-goal-is-curation-not-compilation-llm-wiki-6f90f829b15d — Steven Thompson, "The Goal is Curation, not Compilation"
- Anthropic, "The components of agent memory" (research post, 2026-11-06): https://www.anthropic.com/research/components-of-agent-memory
## Related
- [[Using LLM Wikis to Prepare Courses]]
- [[Andrej Karpathy]]
- [[Large Language Models (LLMs)]]
- [[LLM Knowledge Bases Over Unstructured Data]]
- [[Retrieval-Augmented Generation (RAG)]]
- [[Agentic Knowledge Management (AKM)]]
- [[Personal Knowledge Management (PKM)]]
- [[Obsidian]]
- [[Obsidian Web Clipper]]
- [[Knowledge Graph (KG)]]
- [[Markdown]]
- [[Farzapedia]]
- [[Marp]]
- [[Compounding Knowledge]]
- [[Graph Explorer Base View plugin for Obsidian]]
- [[AI Verifiability]]
- [[AI Verifiability as a Capability Ceiling]]
- [[Menugen Architecture Pattern]]
- [[Markdown-based Installation (MD Scripts)]]
- [[Agent-Native Product Decomposition]]
- [[Open Knowledge Format (OKF)]]
- [[AI Agent Memory]]
- [[Agent Memory Hierarchy]]