In view of the fact that the atmospheric motion is an irreversible process , a memory function which can recall the observation data in the past has been introduced , moreover it is a new method of numerical weather forecast which combine dynamic and statistic . retrospective time integration scheme, which is generalized from the traditional difference scheme can improve the forecast accuracy , it contains historical information so this scheme adapts to mesoscale weather forecast . The purpose of this paper is to apply this scheme to MM5 model and validate the efficiency of this scheme .Based on the atmospheric self-memorization principle ,the retrospective time integration scheme in a mesoscale numerical model is established which is called SMM5 , and the experimental result is compared with the kernel model MM5. the data used in this experiment is taken in the rain season , 1998 and we emphasize on analyze the fields of wind , air pressure ,temperature of 700hpa during 16-28 July 1998 and the field of rainfall was studied too , it shows that: Because of using information of several history fields ,SMM5 can improve the prediction accuracy . verified with both correlation coefficient and root-mean-square error , the wind, air pressure, temperature fields predicted with SMM5 are much better than that of MM5 . As to the rainfall field , both the precipitation areas and precipitation intensities of SMM5 is more similar to the observed field than that of MM5.