Meta's memory agent boosts performance 57% by solving 'behavioral state decay
meta ai introduced a memory agent module that tracks progress and selectively reminds the main agent to prevent repeated errors. this solved 57% more tasks on terminal-bench.
meta ai researchers identified behavioral state decay as the cause of agent failure in long tasks, leading to forgotten constraints and repeated error rediscovery. their solution is a secondary memory agent that reviews a sliding window of steps, updates a structured memory bank, and decides whether to remind the primary action agent about past failures or risks. testing with claude opus as the memory coach and claude sonnet as the worker agent boosted success rates by 57% on terminal-bench and 16% on tau2-bench. the system functions as a plug-and-play layer, requiring no modifications to existing agents.