Meta researchers introduce Proactive Memory Agent to improve AI task performance

Meta researchers have introduced the Proactive Memory Agent (PMA), a system designed to improve the performance of AI agents during complex, long-duration tasks. The research addresses ‘behavioral state decay,’ a phenomenon where an AI agent loses track of requirements, diagnoses, or subgoals as a task progresses.

The PMA architecture functions as a separate entity that sits alongside an unmodified action agent. It maintains a structured memory bank consisting of knowledge, procedures, and private status. By monitoring recent activity, the PMA selectively injects concise reminders into the agent’s context when it detects the agent is drifting off course or repeating mistakes, while remaining silent when no intervention is needed.

In experimental testing, the system demonstrated significant improvements. When paired with Claude Sonnet 4.5, the Terminal-Bench 2.0 pass@1 success rate increased from 37.6% to 45.9%. On the τ²-Bench benchmark, rates rose from 55.0% to 61.8%. Similar improvements were observed using Claude Opus 4.6 and a Qwen3.5-27B memory agent. The researchers noted that selective intervention outperformed both passive memory retrieval and constant, always-active reminders.