Say you're booking a crew to unload a truck. You ask your AI agent when the supplier's coming. It says Friday and pulls up the email. I'd be tempted to go ahead and book it.
A few messages further down the thread, though, the supplier asks to move the delivery to Tuesday. There's your reply. “Tuesday works for me.”
I can just hear the Friday morning phone call. “We're here. Where's the truck?”
How did it get Friday?
RAG stands for retrieval-augmented generation. In this case, the system looks through your emails and gives the model a message to help answer your question. Does that count as memory? Sure. The inbox gives the agent something it can go back to. It just found the old date.
A search could turn up the Tuesday reply too. For a job like this, though, I'd put the change on the job card. “Delivery moved from Friday to Tuesday,” with a link to the email where you agreed. The same search can pull up that card next time.
What did the agent learn?
Saving Tuesday on the job card changes what the agent can find tomorrow. Reading that card doesn't retrain the model. It can use the new date because someone wrote it down where it can find it.
If somebody showed me a demo of an agent with “long-term memory,” that's what I'd try. Give it the original email and the change, then open a fresh chat. “When should I book the crew?” I'd want to hear Tuesday, of course. But before I make that call, I'd ask, “Can you show me where we agreed to that?”
