Knowledge Management
Three articles converging on one finding: retrieval isn’t the hard problem. Memory is. And the best current answer is a compounding LLM-maintained markdown wiki — the Karpathy pattern.
What you’ll find here
- Karpathy’s LLM Wiki pattern — three-layer architecture, three operations (ingest / query / lint), five-project ecosystem
- The orchestrator that was missing — Rezvani’s AutoResearch case for adding decomposition + parallelism above worker agents
- Markdown beat a $50M vector database — the provocative case for simple markdown + good CLAUDE.md over elaborate vector-search infrastructure
Featured in this category
- Why Karpathy’s “LLM Wiki” is the Future — the foundational pattern
- The Orchestrator Was Missing — worker agents need something above them
- The Markdown file that beat a $50M vector DB — start-simple-first
The shift
“We are moving away from treating LLMs purely as search engines or text generators, and finally starting to use them as tireless librarians and system maintainers.”
For Harris / Constellation internal knowledge bases (customer support, onboarding, compliance references), the practical recommendation: start with markdown, add complexity only when you prove you need it.