LLM Knowledge Base

An LLM knowledge base is a structured collection of curated content that large language models can access and reason over directly. Unlike RAG (Retrieval-Augmented Generation), which requires vector databases, embedding pipelines, and chunking strategies, an LLM knowledge base works through structured data formats — Markdown, JSON, or MCP tools — that fit inside a model's context window or are queryable via tool use.

The concept was crystallized by Andrej Karpathy in early 2026 when he described maintaining a personal 'LLM wiki' — a curated set of documents that he feeds to Claude or GPT when he needs deep answers on specific topics. His insight: the bottleneck isn't the model's intelligence, it's the quality of the knowledge you give it. Garbage in, garbage out. A carefully curated 50-article collection beats a 10,000-document RAG index.

For individual knowledge workers, this changes the economics of personal knowledge management. Instead of building elaborate Notion databases or Obsidian vaults that only you can search, you build a collection that both you and your AI agents can query. Burn 451's vault is exactly this: articles you've read and curated, with AI-generated metadata (summaries, key points, relevance scores), exposed through MCP so any AI agent can search your reading history.

The architectural advantage over RAG: no embedding drift, no chunking artifacts, no retrieval failures from semantic mismatches. The tradeoff: it works best at hundreds to low thousands of documents, not millions. For personal knowledge, that's the sweet spot.