Designed agent memory stopped paying off, and the benchmarks missed it
Four papers published in October 2026 found that designed agent memory stopped beating a plain context window with current models, and that the standard benchmarks could not have told you.