interesting study i keep thinking about.
mid-task, the agent still has the whole trajectory in context. it just stops using the part that mattered. nothing was lost. the information is sitting right there, and the agent acts like it isn't.
a july paper called this behavioral state decay. their fix was a separate memory process sitting next to the actor, whose only job is to keep the important state in play. on terminal-bench 2.0 with sonnet 4.5 that took the score from 37.6 to 45.9.
what i like about it is the framing. we keep treating memory as a storage problem: bigger windows, more retrieval, longer transcripts. this says the bottleneck is attention to what's already there. having something and using it are different jobs, and right now we only build for the first one.