New Memory Agent Improves Long-Horizon AI Task Performance

In complex, long-running AI tasks, crucial information from earlier in the trajectory can become buried or lost from the agent's context window, leading to a decline in decision-making quality. This phenomenon, termed "behavioral state decay," hinders the ability of AI agents to effectively complete multi-step objectives. A new approach introduces a "proactive memory agent" designed to actively manage and surface relevant information. This memory agent operates independently but in parallel with an unmodified action agent. Its function is to continuously update a structured memory bank based on recent interactions and then strategically decide when to inject a pertinent reminder into the action agent's context, or to remain silent if no intervention is needed. This plug-and-play module has been tested on challenging benchmarks like Terminal-Bench 2.0 and τ²-Bench, demonstrating notable improvements in task success rates for both weaker and stronger action agents. Ablation studies confirm that this selective, proactive intervention is more effective than passive memory exposure, constant injection, or general retrieval methods. The research also includes initial steps towards developing open-weight memory policies, training a Qwen3.5-27B model to enhance validation reward and achieve partial transferability to other benchmarks.