willynikes2/knowledge-base-server

Implements a seven-layer intelligence pipeline that progressively refines raw documents through classification, promotion, and synthesis into ranked, retrievable knowledge artifacts. Uses three-tier memory (hot/warm/cold) to prevent context pollution while achieving 90%+ token savings through metadata-first retrieval, and feeds captured sessions and fixes back into the system to create a self-modifying instruction loop that compounds agent performance over time. Targets multi-agent workflows via MCP server and REST API, with Obsidian as the human curation layer and SQLite FTS5 + local HuggingFace embeddings for hybrid keyword-semantic search.