QMD: local markdown search and agent memory setup

QMD is a command-line search tool for markdown notes, documentation and meeting transcripts. The video explains how it can give an AI agent access to relevant saved material between sessions. The speaker attributes the project to Shopify CEO Tobi Lütke and describes OpenClaw's use of it as a memory backend.

The search pipeline combines BM25 keyword retrieval through SQLite with vector search and a Qwen reranker. The speaker identifies EmbeddingGemma as the embedding model, with 300 million parameters and a size of about 300 MB. For fuller searches, a fine-tuned 1.7 billion parameter model expands the query into keyword variants, semantic rephrases and hypothetical answers. Folder labels accompany matching sections so an agent receives context about where a document belongs.

A quick keyword probe can skip expansion and vector search when it finds a strong match. QMD also caches results. According to the speaker, its three models download from Hugging Face on first use, totaling about 2 GB, and searches run locally without sending notes or documents away.

The setup outline covers installation, adding a collection, building embeddings and running searches. An MCP server lets Claude Code and other compatible agents query the index during a conversation. The speaker says QMD works on Mac and Linux, while Windows users report embedding hangs and path handling problems.