Karpathy Outlines RAG Alternative — An AI-Curated 'LLM Wiki'

Andrej Karpathy, a co-founder of OpenAI and former Director of AI at Tesla, has outlined a new architecture for managing knowledge in large language models that sidesteps the industry's dominant Retrieval-Augmented Generation (RAG) method. The approach, which Karpathy calls an “LLM Knowledge Base,” uses an AI agent to continuously curate and synthesize information into a single, evolving markdown file, rather than fetching raw documents on demand.

A Curated Alternative to RAG

The standard RAG model works by searching a vector database of documents at the moment a user asks a question, retrieving relevant chunks of text, and feeding them to an LLM as context. This has become the default method for grounding models in proprietary or up-to-date information. Karpathy’s proposal, detailed in a post on X, represents a different philosophy. Instead of a last-minute scramble for information, his system employs an AI to act as a persistent librarian for a specific topic.

As VentureBeat reports, this “LLM Knowledge Base” is a living document. When new information becomes available—such as a research paper, article, or user note—an AI agent reads it, determines its relevance, and integrates a summary into the appropriate section of the markdown file. The knowledge base is therefore constantly being refined, expanded, and de-duplicated, creating a dense, structured summary of a domain. The LLM then uses this curated file as its primary context, bypassing the need for a separate retrieval step at query time.

The 'LLM Wiki' in Practice

To demonstrate the concept, Karpathy shared a GitHub Gist titled “LLM Wiki,” providing a concrete example of what he calls an “idea file.” The code illustrates how the system would function, managing topics and subtopics in a simple text format. The concept quickly gained traction within the developer community, with a corresponding link on Hacker News generating over 250 points and 77 comments, indicating significant interest in alternatives to mainstream RAG pipelines.

This approach moves the complexity away from the query-time retrieval process—which can be brittle and prone to errors—and into a continuous, background curation process. The trade-off is clear: it substitutes the compute cost of real-time search with the cost of running a persistent AI agent that maintains the knowledge base. This is less like a search engine and more like a human expert who reads constantly to stay current in their field, synthesizing knowledge along the way.

The pattern here points to a maturation in how developers are thinking about AI systems. Simply bolting on a document database was a functional first step, but it is not an architecture. Karpathy's proposal treats knowledge integration as a core, stateful process. It suggests a future where AI agents are not just responding to prompts but are actively responsible for maintaining their own understanding of the world, one markdown file at a time. The system's effectiveness will ultimately depend on the quality of the “librarian” AI and whether the economics of continuous curation outweigh the costs of real-time retrieval at scale.