Two different kinds of memory, doing two different jobs
You've already seen agent memory in the context of a single run, remembering facts, past messages, things it looked up. That kind of memory answers questions like 'what did the user say earlier' or 'what did I find when I searched.' Long-horizon tasks need something else entirely, and it's easy to conflate the two if you haven't seen them separated before. What a long-horizon agent needs isn't just memory of facts, it's a saved record of the plan itself: what the overall goal is, which ordered steps from the plan are done, which one is next, and what constraints or decisions were already locked in along the way.
Think back to the marathon coach. The coach doesn't just remember facts about the runner, like their favorite shoes or a joke from last week. The coach keeps a training log: which weeks are complete, what the current mileage target is, and any adjustments already made, like 'switched to trail runs after the knee soreness.' That log is what lets a completely different coach, filling in for one week, pick up the exact same plan without asking the runner to re-explain the whole marathon goal from the beginning. That log is the thing this lesson is about: a durable, external record of a plan's status.