In this paper, a local-learning algorithm for multi-agent is presented based on the fact that individual agent performs local perception and local interaction under group environment. As for individual-learning, agent adopts greedy strategy to maximize its reward when interacting with environment. In group-learning, local interaction takes place between each two agents. A local-learning algorithm to choose and modify agents actions is proposed to improve the traditional learning algorithm, respectively in the situations of zero-sum games and general-sum games with unique equilibrium or multi-equilibrium. And this local-learning algorithm is proved to be convergent and the computation complexity is lower than the NashAdditionally, through grid-game test, it is indicated that by using this local-learning algorithm, the local behaviors of agents can spread to globe.