Karpathy — LLM Wiki Gist

Andrej Karpathy's ~2,000-word "idea file" (April 2026) for a personal knowledge base an LLM builds and maintains: immutable raw/, an LLM-owned wiki/, and a schema file that disciplines the agent. This wiki is an instance of it, and CLAUDE.md descends from it almost line by line (Agent-Maintained Wiki). Agent-fetched 2026-11-08 at the human's request; it had been a #to-read lead for one day.

Summary

The recipe replaces per-query RAG, where "the LLM is rediscovering knowledge from scratch on every question," with a wiki the LLM "incrementally builds and maintains." Three layers (raw sources, wiki, schema), three operations (ingest, query, lint), two navigation files (index.md, log.md), and optional tooling (qmd for search, Marp, Dataview, the graph view). The human curates, directs and asks; the LLM does the bookkeeping, which is why the wiki survives where human-maintained ones are abandoned. It is "intentionally abstract": everything is optional and the agent and human instantiate a version that fits.

Key claims

• The problem: "the LLM is rediscovering knowledge from scratch on every question. There's no accumulation."
• The idea: the LLM "incrementally builds and maintains a persistent wiki … The knowledge is compiled once and then kept current, not re-derived on every query."
• "the wiki is a persistent, compounding artifact. The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything you've read."
• Roles: "Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase." "The human's job is to curate sources, direct the analysis, ask good questions, and think about what it all means. The LLM's job is everything else."
• Ingest: "A single source might touch 10-15 wiki pages. Personally I prefer to ingest sources one at a time and stay involved … But you could also batch-ingest many sources at once with less supervision."
• Query: "good answers can be filed back into the wiki as new pages."
• Scale: index.md read first "works surprisingly well at moderate scale (~100 sources, ~hundreds of pages) and avoids the need for embedding-based RAG infrastructure." Beyond that: "as the wiki grows you want proper search," and he names qmd.
• Log: "if each entry starts with a consistent prefix (e.g. ## [2026-04-02] ingest | Article Title), the log becomes parseable with simple unix tools."
• Why it works: "The tedious part of maintaining a knowledge base is not the reading or the thinking — it's the bookkeeping … Humans abandon wikis because the maintenance burden grows faster than the value."
• Lineage: "related in spirit to Vannevar Bush's Memex (1945) … with the connections between documents as valuable as the documents themselves. The part he couldn't solve was who does the maintenance. The LLM handles that."
• Scope: "This document is intentionally abstract … Everything mentioned above is optional and modular."

What this wiki has since revised (read against Method Reflection)

• "10-15 wiki pages" per source became CLAUDE.md's page quota, and was dropped 2026-11-08 (T5) after Anthropic — The Components of Agent Memory measured bloat. The number was Karpathy's example, not a target.
• "No accumulation" with RAG became CLAUDE.md's "you are not a RAG system," and was qualified (T5): wiki plus search beat either alone, so the wiki is a map on top of search (Navigation vs Search).
• "Stay involved" is Karpathy's own preference, and T4 is what happens when this wiki chose the other branch (unattended auto-ingest) for coverage.
• "The LLM's job is everything else" is T3's premise stated as a design goal, before the question of who ends up holding the knowledge was asked.

Connections

• Agent-Maintained Wiki — the concept this is the primary source for; now triangulated with Anthropic's evaluation and Blumenfeld's scaling critique.
• Method Reflection — the method's origin text; four of its lines map onto T3, T4 and T5.
• Navigation vs Search — the recipe's own scale estimate and its "index first, then search" order.
• Blumenfeld — Scaling Karpathy's LLM Wiki (Neo4j) — the critique; Karpathy's "~100 sources, ~hundreds of pages" is the number Blumenfeld's "can't be scaled" has to meet.
• Bush — As We May Think #to-read — the Memex ancestor: associative trails as the first statement of navigation over documents.
• Dreaming (Offline Memory Consolidation) — "Lint" is the recipe's dream pass.

Source

• raw/2026-04-04 Karpathy — LLM Wiki (gist).md (agent-fetched at the human's request; the gist's single file, verbatim).