A markdown wiki written and maintained by an LLM rather than a human, compiled incrementally from raw source material so that an agent can later query its own structured notes instead of stuffing everything into context.
What it is
An LLM Knowledge Base is a folder of markdown files — articles, concept entries, summaries, backlinks — that an LLM agent compiles from a raw/ directory of source documents and then maintains over time. The agent ingests new material, writes or updates the relevant entries, lints the wiki for inconsistencies, and answers complex queries by reading its own articles. The user typically views the wiki in a markdown editor like Obsidian and rarely edits it directly.
The pattern was popularised by Andrej Karpathy’s April 2026 description of his personal research workflow: ingest, compile, view, query, lint. At small-to-medium scale (Karpathy reports ~100 articles, ~400k words on a single research topic), an LLM agent can answer queries against the wiki without needing a vector database or formal RAG infrastructure — the auto-maintained index files and summaries are sufficient. See the main article for the full workflow.
Why it matters
LLM Knowledge Bases reframe the agent-memory problem. Instead of trying to fit more material into the Context Window or building elaborate RAG pipelines, the agent is given a clean file system and asked to externalise its memory there. Open-source projects like ByteRover, described by Kevin Nguyen, report 50–70% token savings versus dumping documents into prompts because tiered retrieval pulls only the chunks the agent needs.
The pattern is well-suited to bounded domains — a single research project, a case file, a curated topic — where the corpus stays in the thousands of articles rather than the millions. It is inspectable in ways vector stores are not: every claim sits in a markdown file you can read and correct. And it compounds: each query that produces new output gets filed back into the wiki, enriching future queries.