A deterministic memory layer for AI agents that replaces vector similarity search with direct lookup. Instead of embedding chunks and hoping RAG returns the right context, you get exact content retrieval with attribution in under a millisecond. Exposes tools for storing and querying knowledge (grantai_teach, grantai_infer), importing files and git history, and tracking project state across sessions. The multi-agent setup is straightforward: point multiple agents at the same Docker volume and they share memory automatically, with optional speaker attribution if you need to filter by source. Built for compliance scenarios where approximate answers aren't good enough and you're tired of paying retrieval costs on every conversation turn.