LLM Wiki — an "idea file" for building a living knowledge base

What is LLM Wiki?

It has been reported that a public gist outlines a pattern called "LLM Wiki" — an idea file meant to be copy-pasted into your LLM agent (OpenAI Codex, Claude, Pi, etc.). The pitch is simple: stop treating LLMs as one-shot summarizers and instead let them build a persistent, interlinked markdown wiki that sits between you and your raw documents. Think less ephemeral retrieval, more compiled knowledge that improves over time.

How it works

The core contrast is with RAG systems. RAG fetches chunks and recomposes answers on every query — useful, but wasteful. LLM Wiki asks the agent to read new sources, extract the key facts, update entity pages, note contradictions, and revise topic syntheses. The wiki becomes a compounding artifact: cross-links already exist, syntheses reflect everything added so far, and you rarely have to reassemble the puzzle piece by piece. The author suggests pairing the agent with Obsidian as an IDE — the LLM writes, you browse in real time. Neat, right? Relief for anyone who's tired of asking the same five questions and getting five slightly different answers.

Why it matters

Applications are broad: personal journals that actually learn about you, research wikis that evolve into theses, team knowledge bases that don't rot, even fan wikis built chapter-by-chapter as you read Tolkien-style epics. The gist sketches a three-layer architecture — raw immutable sources, the LLM-maintained wiki, and the query/interaction layer — and recommends humans in the loop for oversight. Of course, there are practical caveats: model errors, drift, and trust remain concerns. Still, as agents and personal knowledge tooling converge, this feels less like sci‑fi and more like a sensible next step for anyone who wants their knowledge to compound instead of evaporate.

Sources: gist.github.com, Hacker News