LLM Wiki Pattern
The LLM Wiki pattern is a way to build a persistent, compounding markdown knowledge base with an LLM as the maintainer. Instead of treating sources like a retrieval bucket, the agent reads sources once, preserves them in raw-source-ingestion, and integrates their important claims into durable wiki pages.
Core idea
Traditional RAG retrieves chunks at query time and reconstructs knowledge repeatedly. The LLM Wiki pattern compiles knowledge into an interlinked wiki so synthesis, contradictions, and relationships accumulate over time.
For Dagobert, this means Johannes can ask for research on any recurring topic and Hermes should add it to this wiki rather than letting the result disappear into chat history.
Roles
- Johannes curates directions, questions, and source priorities.
- Hermes captures sources, summarizes, cross-references, maintains
index.md, and logs changes. - Dagobert hosts the markdown/wiki files under
/research/and can later expose HTML dashboards or summaries.
Practical rules for this wiki
- Preserve original source captures in
raw/. - Create pages for central concepts/entities only.
- Keep pages interlinked with
[[wikilinks]]. - Update pages when new sources change the current synthesis.
- Store substantial answers in
queries/so future work can build on them.