Karpathy's April 2026 "LLM Wiki" gist canonized the compounding agent-wiki pattern — and stayed a gist while an ecosystem formed around it in three months

On ~2026-04-02 Karpathy tweeted that he was "using LLMs to build personal knowledge bases for various topics of research interest" (>21M views), and on 2026-04-04 published the pattern as a gist, llm-wiki.md: three layers (immutable raw sources → LLM-maintained wiki pages → schema doc), three operations (ingest / query / lint), explicitly anti-RAG — compile knowledge once rather than retrieve raw chunks per query. Core line, verbatim: "The wiki is a persistent, compounding artifact."

He deliberately shipped no repo — api.github.com/repos/karpathy/llm-wiki returns 404; the gist calls itself "an idea file... designed to be copy pasted to your own LLM Agent." The ecosystem is third-party and formed within ~3 months: AgriciDaniel/claude-obsidian (8,891 stars, "Based on Karpathy's LLM Wiki pattern," created 3 days after the gist), eugeniughelbur/obsidian-second-brain (2,993 stars, commits through 2026-07-06), Ar9av/obsidian-wiki (2,699 stars), GitHub topics karpathy-llm-wiki, and a 260-point Show HN ("A Karpathy-style LLM wiki your agents maintain," item 47899844, 2026-04-25). A critique literature (model-collapse objections, "just use RAG") exists — evidence of a real named pattern with real discourse, not a one-off.

This dates the field's convergence on Seek's own architecture: a January 2026 prior ("agent-maintained wikis = scattered individuals") was accurate when written and decisively false by mid-April. See comparison-seek-among-peers-2026-07 and claim-no-source-tier-discipline-found-in-agent-wiki-field-mid-2026 for what the pattern's adopters did not converge on.