The idea that a language model can be your research partner is not new. But using one to build a self-organizing, cross-linked knowledge base that improves as you feed it more information — that's a shift worth paying attention to.
The concept involves feeding raw source documents — transcripts, articles, PDFs — to a coding agent, which converts them into structured markdown files complete with summaries, backlinks, concept tags, and relationships to other entries. The result is a queryable wiki that grows smarter as your collection expands. The approach was recently popularized by AI researcher Andrej Karpathy, whose writeup on the method gained significant traction almost immediately.
Why Markdown Is the Right Foundation
The choice of markdown as the storage format isn't arbitrary. Markdown files are lightweight, version-controllable, and natively readable by LLMs. When an agent processes your source material into markdown, it's creating a format it can later query directly. Tools like Obsidian make the file graph visually navigable, but the intelligence lives in the files themselves — not in a proprietary database.
The workflow is straightforward: drop source documents into a folder, instruct your agent to process them into wiki entries, and let it handle relationship-building automatically. The agent identifies shared concepts, creates backlinks between entries, and surfaces patterns you might miss manually. Someone who fed 36 video transcripts into this system found the agent independently mapped connections between tools, frameworks, and techniques — without any manual tagging.
Compounding Value Over Time
The core advantage over a regular note-taking system is compound growth. Each new document doesn't just add one entry — it adds potential connections to every existing entry. An article on retrieval pipelines might automatically link to six previous entries on embedding models, search strategies, and context management.
Traditional semantic search RAG requires embedding infrastructure, chunking strategies, and vector databases. The markdown wiki approach sidesteps this complexity. Once the wiki contains a hundred or more entries, you can query it conversationally — asking things like "what trade-offs have I encountered between X and Y?" — and the LLM synthesizes answers from the structured file graph.
Separate Wikis for Separate Domains
The approach scales well when you treat different knowledge domains as separate wikis. A technical research wiki can sit alongside a business operations wiki containing quarterly goals, team context, and project history. These can remain isolated or be merged depending on what context a downstream agent needs, without bloating any single knowledge store.
Takeaway
LLM-powered knowledge bases represent a practical alternative to elaborate RAG pipelines for personal or small-team research. The entry cost is low — a folder of documents, a markdown viewer, and a capable coding agent — but the long-term payoff is a system that compounds in value as you use it. If you're building any research or content workflow, this is the architectural pattern to experiment with first.