The LLM-Powered Wiki: A Persistent Alternative to Throwaway RAG
The core critique of standard RAG is that it is stateless by design — every query triggers a fresh retrieval and synthesis pass with no memory of prior work. For complex questions requiring synthesis across multiple documents, this means the LLM is repeatedly reassembling the same puzzle from scratch. The proposed LLM-powered wiki pattern addresses this by front-loading the synthesis work into a persistent, structured knowledge base.
The architecture has three layers: immutable raw sources the LLM reads but never modifies, a wiki directory of markdown files the LLM owns and maintains entirely, and a schema/config file defining how the wiki is structured. Three core operations drive the system — ingest (process a source, extract concepts, update and cross-link existing pages, flag contradictions), query (answer questions from the wiki and file useful answers back as new pages), and lint (maintenance pass to catch stale claims, orphan pages, and missing cross-references). A single ingested source may touch 10–15 wiki pages.
A key design argument is that humans fail at wikis because maintenance cost grows faster than value — people forget to update cross-references and eventually abandon the project. LLMs, by contrast, can touch 15 files in a single pass without fatigue, collapsing the maintenance burden to near zero. The wiki is also designed to be portable: local markdown files, no vendor lock-in, compatible with any tool, and swappable between AI providers including open-source models or a fine-tuned model trained on the wiki itself.
The host demonstrated the system by building it in Claude Code with eight trading strategy transcripts as raw sources, completing a basic working version in approximately one hour. Two agents ran in parallel during initial ingestion, and the system performed live web searches to backfill the wiki when a query surfaced gaps in existing coverage — behavior the host framed as the intended growth loop.
One notable caveat: the host referenced "Opus 4.6" as the Claude model used, which does not match any publicly released Claude version, raising minor questions about the accuracy of model identification in the demo. The broader architecture claims remain unverified but are internally consistent and practically testable.
Standard RAG forces LLMs to rediscover knowledge from scratch on every query with no accumulated understanding. The proposed alternative: let an LLM build and maintain a persistent, cross-linked markdown wiki up front, so synthesis and context are already there when you ask questions. The host built a working version in roughly an hour using Claude.
The wiki is explicit (all visible in navigable markdown), yours (local files, no provider lock-in), file-over-app (universal formats, interoperable with any tool), and bring-your-own-AI (plug in Claude, GPT, open-source models, or fine-tune a model on your wiki).
In standard RAG, nothing accumulates. Every time you ask a question, the LLM is rediscovering knowledge from scratch, repiecing together fragments every single time.
If you ask something subtle that requires synthesizing five different documents, it has to find and connect all those pieces on every query. There's no memory, no cross-references, no accumulated understanding.
You never write the wiki yourself. The LM writes and maintains all of it.
A single ingested source might touch 10 to 15 wiki pages.
When ingesting a source, the LLM executes seven steps: reads raw source, extracts key concepts/entities/claims, writes a summary page, updates existing entity and concept pages, flags contradictions, updates the index, and appends to the log.
Humans abandon wikis because the maintenance burden grows faster than the value.
LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in a single pass, dropping the cost of wiki maintenance to near zero.
You can fine-tune a model on your wiki so it knows your data in its weights, not just in its context.
When a query surfaces information not already in the wiki, the agent performs a web search and automatically backfills the wiki with the new information found.
The system can also be used with OpenClaw or Hermes agent, not just Claude.
The host used 'Opus 4.6' to describe the Claude model used, which does not correspond to any publicly released Claude model version as of the knowledge cutoff, raising questions about accuracy of model identification.
The LLM-powered wiki pattern flips standard RAG: instead of retrieving at query time, the LLM builds a persistent wiki up front, with cross-references already there, contradictions already flagged, and synthesis already reflecting everything fed to it.
The architecture has three layers: raw sources (immutable, LLM reads but never touches), the wiki itself (a directory of markdown files the LLM owns entirely), and the schema (a configuration file like a Claude MD that tells the LLM how the wiki is structured).
There are three core operations: ingest (drop a source, LLM processes and cross-links), query (ask questions, synthesize from wiki, file answers back), and lint (maintenance pass to find contradictions, stale claims, orphan pages, missing cross-references).
Good query answers can be filed back into the wiki as new pages, so explorations compound in the knowledge base just like ingested sources do.
The host built the LLM wiki system in Claude Code using Opus 4.6, starting with eight trading strategy transcript files as raw sources, and completing the basic version in approximately one hour.
Claude ran two agents simultaneously to ingest raw sources, create summaries, and update concept pages during the initial build.
“The LM incrementally builds and maintains a persistent wiki, structured interlinked markdown files sitting between you and your raw sources.”
“This is exactly how the wiki is meant to grow. You ask a question, I research beyond the wiki, and the new knowledge got filed back as permanent pages.”
Shared with Earmark · earmark-ai.com