LLM Wiki Pattern — Founding Idea File

Source: raw/llm-wiki-pattern.md Ingested: 2026-04-16 Author: Unknown (shared by user as founding idea document)

---

Summary

A pattern for building a personal knowledge base where an LLM incrementally builds and maintains a persistent wiki, rather than re-deriving knowledge from raw documents on every query (RAG). The wiki is a compounding artifact that gets richer over time. The human curates and directs; the LLM does all the bookkeeping.

Key Takeaways

1. RAG re-derives, wikis compound. RAG re-discovers knowledge from scratch per query. A persistent wiki accumulates synthesis — cross-references, contradiction flags, and summaries are computed once and kept current.
2. Three layers: raw sources (immutable), the wiki (LLM-owned), and the schema (CLAUDE.md — the operating rules for the LLM).
3. Three operations: ingest (add a source), query (ask a question), lint (health-check the wiki).
4. Index + log duality: index.md is content-oriented (what exists); log.md is time-oriented (what happened).
5. LLM solves the maintenance problem. Humans abandon wikis because upkeep costs exceed value. LLMs don’t get bored and can touch 15 files in one pass.

Concepts Introduced

• LLM Wiki Pattern — the core approach described in this document
• RAG vs Wiki — implicit contrast between retrieval-augmented generation and persistent wiki

Intellectual Lineage

• Vannevar Bush’s Memex (1945): private, associatively-linked knowledge store. Bush’s unsolved problem was who maintains it. This pattern’s answer: the LLM.

Open Questions

• How does the approach scale beyond ~100 sources / hundreds of pages without embedding-based search?
• What’s the right grain for entity vs. concept pages in a personal context?
• How to handle sources that contradict each other heavily?