The evidence chain is the part I keep coming back to. Most agent memory is just “vector DB + hope for the best.” Having structured provenance where you can trace a claim back to the actual email that confirmed it is a game changer for trust in agent-operated systems.

Worth noting this is intentionally headless — no UI shipped. The bet is that between MCP and AI code gen tools, the bottleneck is no longer the frontend, it’s having a solid typed API behind it. I’ve been using it with Claude Desktop and the conversational interface is surprisingly natural for CRM workflows.