Agent Memory: How Does It Find Anything? (Ep 4)

​Retrieval is where most teams are actively tuning, and it's also where added sophistication quietly stops paying — adding a planning step to retrieval improved every metric tested, but adding a reflection step on top of that planning made things measurably worse.

​The deeper finding is that surfacing one relevant memory first and assembling complete evidence are different problems with different winners, so the system that takes Recall@1 isn't the system that takes Recall@10. We work through when to optimize for which.