Every AI app claims it can “remember” you. Most of them don’t.

When I built my voice AI agent, I watched it forget my salary, hallucinate my daughter’s name, and confidently describe a restaurant I never visited. So I stopped trusting the usual memory pattern: summarize the chat, store embeddings, and hope retrieval works later.

In this talk, I’ll walk through the three-layer memory architecture I built after those failures: a deterministic profile, a facts table, and vector search over transcript chunks. I’ll show where the first version broke, why whole-transcript embeddings made recall worse, and why the model should propose memory but never own truth.

You’ll leave with a practical way to design AI memory that survives corrections, voice latency, long-term context, and the boring facts users expect your app to get right.