Concept — LLM Wiki Pattern
A pattern for building a personal knowledge base with an LLM: the LLM incrementally builds and maintains a persistent, interlinked wiki that sits between the human and the raw sources.
Definition
The traditional approach (RAG) retrieves relevant snippets from raw at query time and generates an answer — rediscovering knowledge from scratch each time, with no accumulation. The LLM Wiki has the LLM read new sources, extract key information, and integrate it into the existing wiki (updating entity pages, revising topics, flagging contradictions), so that knowledge is “compiled once and kept continuously up to date.”
Key Differences from RAG
| Dimension | RAG | LLM Wiki |
|---|---|---|
| Knowledge location | Re-fetched from raw at query time | Persisted in wiki pages |
| Cross-references | Re-assembled each time | Already present |
| Contradiction flags | None | Already flagged |
| Synthesis | Recomputed each time | Reflects everything read |
| Compounding | No | Yes |
The wiki is a compounding artifact: every source added and every question asked makes it richer. See knowledge-compounding.
Division of Labor
- Human: curate sources, guide analysis, ask good questions, think about “what does all this mean.”
- LLM: summarization, cross-referencing, archiving, bookkeeping — all the drudgery.
A practical picture: an LLM Agent open on one side, Obsidian on the other, watching links update in real time, viewing the graph view, reading updated pages. “Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase.”
Scope
Personal growth, deep research dives, reading-companion wikis, internal team wikis, competitive analysis, due diligence, trip planning, class notes, hobby deep-dives — anywhere knowledge accumulates over the long term and you want to organize it.
Related
- three-layer-architecture — the structure of the pattern
- three-operations — ingest / query / lint
- 2026-karpathy-llm-wiki — original definition
- andrej-karpathy — originator