Agent Memory: What information is worth remembering? (Ep 3)

​Most teams pour their engineering effort into the extraction prompt, tuning what gets distilled out of a conversation before it's ever written down, which the research suggests is exactly backwards. Switching one system from fast to fine-grained memorization collapsed multi-step reasoning tenfold while buying almost nothing in factual retrieval, and cruder heuristic segmentation beat LLM-driven segmentation outright.

​We look at why filtering at write time destroys context you can't recover, and what it costs to defer that decision until read time instead.