Knowledge Systems

Wiki-LLM — A Personal Knowledge Graph

Read once. Know forever.

A method for maintaining a structured, interconnected knowledge base with an AI agent. Rather than the usual retrieval pattern, where every question starts from scratch, this builds a persistent wiki that accumulates and synthesises over time.

The problem it addresses is familiar to anyone who has spent a month uploading documents to a chat window: nothing accumulates. Five documents in, you are still asking the model to re-derive connections it already found last week.

Three layers

• Raw sources. The original documents — PDFs, reports, research, notes. Immutable, and read-only to the agent.
• The wiki. Markdown pages in folders: entity pages, concept pages, source summaries and synthesis pages. Read in Obsidian, with graph visualisation and backlinks.
• The schema. A CLAUDE.md documenting the organisational conventions and the ingestion, query and maintenance procedures — refined as the system teaches you what actually works.

How it runs

• 1. New sources go into a designated folder.
• 2. Claude reads them and discusses what's in them.
• 3. Wiki pages are created or updated, cross-referenced with [[wikilinks]].
• 4. Index and changelog files are kept current.
• 5. Everything is committed to version control.
• 6. Later questions draw on the synthesised wiki rather than the raw sources.

Why it beats chatting with PDFs

The wiki is a compounding asset. Cross-references persist, contradictions get flagged rather than forgotten, and connections found months ago are still there — so an answer can be informed by the whole knowledge base instead of whichever documents happen to be in the current conversation.

Builds on the LLM Wiki pattern introduced by Andrej Karpathy.

← Back to other work