Abstract: In Multi-Agent System（MAS），negotiation between agents is a complicated process in which negotiation agents mutually exchange offers.How to improve the negotiation efficiency between agents has become the focus on which the researchers pays attention.The paper proposes Agent negotiation algorithm which introduces the reinforcement learning idea based on memory-evolution theory.Compared with standard reinforcement learning，the negotiation algorithm included three stage memory-evolution reinforcement learning，which makes agent weigh between real-time response and delay-time response，and creats the interaction condition of social memory for Agent，that makes reinforcement learning fitter MAS requirement.Finally，the paper proves that negotiation algorithm is efficient by simulation experiment.

Key words: memory-evolution, negotiation algorithm, reinforcement leaning

摘要： 在多Agent系统（MAS）环境中，协商是一个复杂的动态交互过程。如何提高协商效率，成为了研究者关注的焦点。应用记忆演化理论的强化学习思想，提出一种Agent协商算法。它与基本强化学习相比，3阶段的记忆演化的强化学习，使得Agent可以在实时回报与延迟回报间更好的做出平衡，并为Agent记忆社会化交互创造条件，使强化学习更适合MAS的要求。通过模拟实验证明该协商算法是有效性的。

关键词: 记忆演化, 协商算法, 强化学习