Former Tesla Director of AI and OpenAI co-founder Andrej Karpathy has shared details of a new approach to AI knowledge management that bypasses traditional vector database systems in favor of AI-maintained Markdown files.

The system, which Karpathy calls "LLM Knowledge Bases," works in three stages. Raw research materials including papers, repositories, and web articles are first collected in a directory. The AI then "compiles" this data by reading through it and creating structured wiki-style summaries with cross-references between related concepts. Finally, the system runs maintenance checks where the AI scans for inconsistencies and updates connections between ideas.

Karpathy's approach differs from the dominant Retrieval-Augmented Generation (RAG) method, which converts documents into mathematical vectors stored in specialized databases. Instead, his system relies on human-readable Markdown files that can be directly inspected and edited. At a scale of approximately 100 articles and 400,000 words, Karpathy reports the system performs effectively without requiring vector database infrastructure.

The methodology has drawn attention from AI researchers and entrepreneurs who see potential enterprise applications. Community members have noted that many companies struggle with unstructured internal data across platforms like Slack and internal wikis that could benefit from automated compilation and organization.

Several variations of the approach have emerged, including multi-agent systems that use quality control mechanisms to validate information before it enters the knowledge base. These implementations aim to address concerns about AI hallucinations potentially corrupting the collective knowledge store.

Karpathy suggests the system could eventually enable fine-tuning smaller AI models on the curated knowledge base, creating specialized intelligence systems trained on specific organizational or research domains. The approach represents a shift toward treating AI as an active maintainer of its own memory rather than a stateless tool that requires context reconstruction for each session.