How do you handle memory in AI agents?

I split memory into a few layers. Working memory is what sits in the context window right now: recent messages, the current plan and tool results. When it grows too large, I keep the last several turns verbatim and replace older ones with a running summary, so the important decisions survive without the full transcript.

Long-term memory lives outside the model. Some of it is structured, such as a user's timezone, preferred language or the status of an open ticket, and that belongs in a normal database with clear fields. Some of it is unstructured, such as notes from past sessions, which I store as text with embeddings and retrieve by relevance and recency when a new request comes in. Only the few relevant memories go into the prompt.

The design questions are mostly about policy. What is worth saving? I prefer explicit, durable facts over saving everything. How do we handle conflicts and stale facts? Newer facts win and old ones expire. And privacy: memory is scoped per user, users can see and delete it, and sensitive data is not stored without a reason.