Researchers from Carnegie Mellon University (CMU) and the Massachusetts Institute of Technology (MIT) have introduced a novel approach called Agent Workflow Memory to enhance the performance of LLM-based models in handling long-horizon tasks with complex action trajectories. This method involves inducing commonly reused workflows from agent experiences and providing these workflows to the agent on demand. The approach, which includes a system prompt, works both offline and online and has shown significant improvements in task success rates. The project, led by Z Z Wang, J Mao, D Fried, and G Neubig, is generating excitement in the AI research community. The research has been documented under the identifier 2409.07429.