Meta Memory Agent Raises Long-Task Benchmark Scores for AI Systems
The Proactive Memory Agent selectively reminds AI systems of important requirements during extended tasks, but the results do not establish consumer-product performance.
Meta researchers proposed a Proactive Memory Agent that raised results on long-horizon AI benchmarks by selectively reminding systems about requirements, diagnoses and subgoals that might otherwise fade during extended tasks.
The research paper, published on arXiv July 10, 2026, describes behavioral state decay as a failure mode in which information an AI agent has learned stops influencing later decisions. PMA is designed to address that problem without changing the separate agent responsible for taking actions.
PMA reviews recent activity, maintains a structured memory bank covering knowledge, procedures and private status, and decides whether to send a short reminder. It can also remain silent when intervention is unnecessary.
With Claude Sonnet 4.5 serving as the action agent and Claude Opus 4.6 handling memory, Terminal-Bench 2.0 pass@1 increased to 45.9% from 37.6%, an 8.3-percentage-point gain. The task-weighted τ²-Bench average rose to 61.8% from 55.0%.
When Claude Opus 4.6 handled both roles, Terminal-Bench 2.0 increased to 45.9% from 43.5%, while τ²-Bench rose to 68.7% from 66.2%.
A Qwen3.5-27B memory agent trained with supervised fine-tuning and reinforcement learning also improved results. Paired with a frozen Qwen3.5-122B-A10B action agent, it raised Terminal-Bench 2.0 pass@1 to 41.1% from 37.6% in an 85-task evaluation.
Selective intervention outperformed passive exposure to the full memory bank and general memory retrieval in ablation tests. The findings apply to the tested models, benchmarks and evaluation setups and do not establish performance in consumer products or production deployments.
The work is relevant to Meta’s broader push into personal AI agents. Meta introduced Muse on Sept. 8 as a personal AI agent that can pursue goals, work across connected applications and continue tasks after users close the app. Its announcement does not identify PMA as part of Muse’s architecture.
Meta’s Muse rollout in the United States describes an agent built to plan and complete tasks across connected services. The PMA research instead focuses on preserving useful task state during long sequences of AI actions.
