A bit more on Cortex since that's the part most people will be curious about:
The retrieval shape is scoped pgvector cosine over ai_memories and org_brain, joined with the entity graph so a single query like "what's the latest with YC?" pulls every memory, every cortex_event, every CRM record, every email thread, every meeting tied to that entity — across modules, deduplicated, in one shot.
The heartbeat aggregator runs on a schedule and rolls fresh cortex_events into memories: pattern detection, strategic shifts, recurring asks. So the AI doesn't just have raw event logs, it has summarized intent.
Retrieval threshold is per-tool, not global, because confidence requirements for "draft me an email" vs "fire a SAFE" are very different.
All of it is exposed to Claude through MCP. You can plug your own Claude into your StartupStarter workspace and get the same context the in-product S2X assistant has.
The retrieval shape is scoped pgvector cosine over ai_memories and org_brain, joined with the entity graph so a single query like "what's the latest with YC?" pulls every memory, every cortex_event, every CRM record, every email thread, every meeting tied to that entity — across modules, deduplicated, in one shot.
The heartbeat aggregator runs on a schedule and rolls fresh cortex_events into memories: pattern detection, strategic shifts, recurring asks. So the AI doesn't just have raw event logs, it has summarized intent.
Retrieval threshold is per-tool, not global, because confidence requirements for "draft me an email" vs "fire a SAFE" are very different.
All of it is exposed to Claude through MCP. You can plug your own Claude into your StartupStarter workspace and get the same context the in-product S2X assistant has.