Developer Builds Durable AI Knowledge Base Using Markdown, Git, and Agent Rules

A software developer has shared a structured approach to building a long-term AI knowledge base using a tiered directory system, Markdown files, and Git for version control. The system organizes information across nine folder types — from raw captures and source records to syntheses, project logs, and archives — each governed by distinct rules about how content can be changed. Unlike standard retrieval-augmented generation (RAG) tools, which reassemble answers from document chunks at query time, this approach maintains a persistent, interlinked wiki that agents actively update with new evidence, cross-references, and noted contradictions. The design draws partial inspiration from a April 2026 proposal by AI researcher Andrej Karpathy, who suggested placing a maintained wiki between raw sources and users. The developer treats RAG and the wiki model as complementary, using search for navigation while keeping curated knowledge and raw evidence in separate, auditable layers.