• title: LLM Wiki Patterncreated: 2026-06-01updated: 2026-06-01type: concepttags: [pattern, knowledge-base, research]sources: [raw/articles/karpathy-llm-wiki-2026.md]
• LLM Wiki Pattern
  • Core Idea
  • Three Layers
  • Division of Labor
  • Applicable Contexts
  • Implementation Notes
  • Relevance to Pluronymous

§title: LLM Wiki Pattern created: 2026-06-01 updated: 2026-06-01 type: concept tags: [pattern, knowledge-base, research] sources: [raw/articles/karpathy-llm-wiki-2026.md]

§LLM Wiki Pattern

Andrej Karpathy’s pattern for building persistent, compounding knowledge bases using LLMs. Based on his gist: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f

§Core Idea

Traditional RAG: upload documents → LLM retrieves chunks at query time → generates answer. Knowledge is rediscovered from scratch every time. Nothing accumulates.

LLM Wiki pattern: the LLM incrementally builds and maintains a persistent wiki — structured, interlinked markdown files. Knowledge is compiled once and kept current, not re-derived per query.

Key difference: the wiki is a persistent, compounding artifact. Cross-references are already there. Contradictions are already flagged. Synthesis already reflects everything ingested.

§Three Layers

1. Raw sources — immutable. Agent reads but never modifies.
2. The wiki — agent-owned markdown files. Created, updated, cross-referenced.
3. The schema — SCHEMA.md defines structure, conventions, tag taxonomy.

§Division of Labor

• Human: curates sources, directs analysis, asks the right questions
• Agent: summarizes, cross-references, files, maintains consistency
• Obsidian (optional): the IDE where you browse the wiki; the LLM is the programmer; the wiki is the codebase

§Applicable Contexts

• Personal (goals, health, psychology)
• Research (going deep on a topic over weeks/months)
• Reading a book (chapter-by-chapter companion wiki)
• Business/team (internal wiki from Slack, meetings, docs)
• Competitive analysis, due diligence, trip planning, hobby deep-dives

§Implementation Notes

• File names: lowercase, hyphens, no spaces
• YAML frontmatter on every page (title, dates, type, tags, sources)
• [[wikilinks]] for cross-references (min 2 per page)
• Append-only log
• Agent orients itself every session: read SCHEMA → read index → scan recent log
• Never modify raw sources — corrections go in wiki pages
• Handle contradictions explicitly with provenance markers

§Relevance to Pluronymous

This IS the pattern we’re using for /home/wiki/. The wiki is our persistent knowledge base for the entire project — architecture decisions, technology research, agent design patterns, deployment runbooks, and cross-cutting concerns that don’t belong in any single crate’s README. ¹