Concepts Sources Graph Archive Round Robin →

Karpathy LLM Wiki

Concept·1 cited sources·updated 2026-04-27

A pattern for personal knowledge bases in which an LLM agent incrementally builds and maintains a structured, interlinked wiki sitting between the user and the raw sources.

What it means

A Karpathy LLM Wiki is not a retrieval-augmented chat over a pile of files. It is a persistent, structured artifact that the LLM compiles once and keeps current — a directory of interlinked markdown pages organized by entity type, with cross-references, contradictions flagged, and a schema file that tells the LLM how to maintain it. The user curates sources and asks questions; the LLM does all the bookkeeping.

The pattern is deliberately abstract — it specifies a shape (three layers, three operations) and leaves implementation specifics (directory layout, frontmatter, slug rules, page formats) to be co-evolved between the user and the LLM agent for the user's domain.

How it shows up in sources

  • LLM Wiki — the canonical statement of the pattern. Defines the three layers, three operations, and the principle that the wiki is "a persistent, compounding artifact."

Mechanism / how it works

The pattern has three structural elements:

  1. A three-layer architecture — see Raw immutable layer, Wiki curated layer, Schema as program.
  2. Three operations — see Ingest, query, lint operations.
  3. Two navigational primitivesindex.md (content-oriented catalog) and log.md (append-only chronological journal). At moderate scale (~100 sources, hundreds of pages) these replace the need for embedding-based RAG.

The compounding behavior comes from a fourth, derived property: maintenance cost approaches zero (the LLM does the bookkeeping that historically killed human-maintained wikis), so the substrate gets structurally richer with every ingest rather than decaying. See Self-improving substrate.

Related concepts

Related vendors / sectors

(this concept is pre-domain — it does not yet wiki-link to specific vendors or sectors. The whole point of this vault is to instantiate the pattern in the digital-asset vendor research domain; that instantiation appears across every vendor / sector / source page.)

Open questions

  • At what corpus size does index-only navigation break down? Karpathy cites "moderate scale (~100 sources, ~hundreds of pages)" as the regime where this works without a search tool.
  • For domains with proprietary mixed-source corpora (e.g. 51 Terminal's audit-context research), how does the pattern adapt to enforce citation-and-licensing discipline structurally? (This vault's Path A licensing posture is one answer.)

Sources cited