URL_033: Karpathy (2026) — 'LLM wiki', the LLM-curated knowledge-base pattern the Neobiome wiki follows

URL_033: Karpathy (2026) — 'LLM wiki', the LLM-curated knowledge-base pattern the Neobiome wiki follows

Source

https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f — original source (opens in a new tab; the file is not redistributed)

Context: thesis · No NI feeds (shallow ingest)

Summary

Karpathy’s “LLM wiki” gist sets out a pattern for building a personal knowledge base with an LLM that goes beyond retrieval-augmented generation (RAG). Where RAG re-retrieves raw chunks and “rediscovers knowledge from scratch on every question”, the LLM-wiki approach has the model incrementally build and maintain a persistent, interlinked set of markdown pages — “a persistent, compounding artifact” in which “the cross-references are already there” and “the synthesis already reflects everything you’ve read”. The design separates three layers: raw sources (documents the LLM reads but never mutates), the wiki (LLM-maintained summary/entity/cross-reference pages), and the schema (a configuration document defining the wiki’s structure, conventions and workflows), operated through three actions — ingest (turn new sources into wiki pages), query (answer from the synthesised wiki rather than raw documents), and lint (keep it consistent, flag contradictions, surface gaps). It is the direct, credited origin of the Neobiome project’s knowledge system: the wiki/ here implements exactly these three layers and three operations. For the project it is the contemporary pattern its method follows, and it stands as the modern descendant of Vannevar Bush’s 1945 “memex” (LIT_116).

Key thesis insights

• The three-layer architecture (raw / wiki / schema). Raw sources are immutable inputs the LLM reads but never edits; the wiki is the LLM-curated layer of summary, entity and cross-reference pages; the schema is a single configuration document that tells the LLM how the wiki is structured and maintained. The Neobiome corpus maps one-to-one: source-files/ + primary-research/ (immutable raws), wiki/ (LLM-maintained pages), and the CLAUDE.md files + controls/ config as the schema. URL_033
• Three operations — ingest, query, lint — and the “compounding artifact” claim. The pattern’s advance over RAG is that synthesis is done once at ingest and reused, so the knowledge base compounds rather than being rebuilt per query: “the LLM is rediscovering knowledge from scratch on every question” is the failure mode it avoids. The Neobiome workflow instantiates all three (the /ingest_source + batch_ingest flow, query-time reading of wiki/index.md, and lint_wiki.py). URL_033
• Lineage. Karpathy’s LLM-wiki is the modern, LLM-native realisation of an old idea — Bush’s memex (1945): a personal store of one’s records, mechanised for associative cross-referencing. Citing the two together traces the method from the memex vision to its LLM implementation. LIT_116

Notes

• Provenance. Verbatim gist markdown retrieved from the GitHub raw endpoint (gist.githubusercontent.com/karpathy/442a6bf555914893e9891c11519de94f/raw, curl). First-party authoritative artifact (the gist itself is the thing being cited), not AI-prepared — no ## Retrieval provenance block required.
• Identity confirmed against the raw. Author Andrej Karpathy (@karpathy); gist file llm-wiki.md; title “LLM Wiki: A pattern for building personal knowledge bases using LLMs”; created 4 April 2026; the three-layer / three-operation architecture as summarised above.
• Relationship to the held concept doc. The full gist text (with Neobiome relevance commentary) has been held in the project’s internal notes since 2026-05-27; this page formalises it as the citable wiki source rather than a loose reference.
• Scope. context: thesis, feeds: [], shallow ingest — no key_claims, no NZ data, no NI calculation-cell exposure. No research target opened or resolved.

Connections

Links to

Sources (1): LIT_116

Referenced by

Sources (1): LIT_116