Your AI has amnesia. Every chat starts from zero, and every tool, whether Claude, ChatGPT or Copilot, remembers something different, or nothing. The obvious way out is a memory of your own, outside any tool. But two owners live there: what's yours, and what belongs to your employer, which isn't yours.

The hippocampus is what turns what you lived today into long-term memories; without it, amnesia. In this session we build one for your AI, with the same T-SQL on two engines: SQL Server 2025 on-premises for the personal side and Azure SQL in the cloud for work. Chats come in through MCP ("save this") or by export, get chunked and vectorized inside the database (AI_GENERATE_CHUNKS, AI_GENERATE_EMBEDDINGS), are retrieved by similarity over the VECTOR type, and every query and every crossing between hemispheres is recorded in a ledger table that nobody can rewrite without it showing. Live, we try to jump the wall, watch it get refused, and verify the record on both sides.

We also test the tempting alternative, a single database with row-level security, and see why it falls short when the data has two owners. You'll leave with the schema, the T-SQL and a clear rule: RLS separates users, not owners.