Memory that makes people come back

Memory is what an agent knows at the start of a session that it did not learn in that session. Working memory is the context window, gone when the call ends. Long-term memory is data you persist and retrieve: episodic memory records what happened, semantic memory records facts and preferences, procedural memory records how to do things. A production memory system is mostly policy, not storage. A write policy decides what may be remembered, from which source, and whether a human must confirm it. A retrieval policy filters by scope first, then ranks by relevance, recency and importance. Update rules resolve conflicts without silently overwriting. Forgetting rules expire, decay and delete, including on an erasure request. And memory is an attack surface: one injected "fact" written today is replayed into every future session.

Key takeaways:

• Memory is not workflow state. "Where is case PA-88213?" belongs to the durable workflow (Workflows that survive restarts); "this provider office prefers fax" belongs to memory.
• Write less than you think. An allow-listed schema of things worth remembering beats "extract everything and let retrieval sort it out".
• Scope is a hard filter in code, never a retrieval score or a prompt instruction.
• Every memory needs provenance, a confidence, an expiry and a deletion path before it is written.
• Memory written from untrusted input stays untrusted on the way back in.