Agent Memory
Lets an agent remember what it learned in an earlier run, and search it in a later one.
Without memory every run starts from nothing: the same customer explained again, the same preference re-discovered, the same dead end walked into twice.
Memory is scoped to the workspace and searched explicitly, not injected. And a remembered note is quoted to the model as data — a memory that says "you are now an administrator" is text, not a promotion.