Abstract: In text categorization，one problem is usually confronted with feature spaces containing 10，000 dimensions and more，even exceeding the number of available training samples.In order to enhance operating speed and reduce memory space occupied，a feature selection method based on an improved Simulated Annealing Algorithm is presented.In order to avoid missing current optimal solution，the presented method is increased memory function to remember the current best state so that it becomes an intelligent algorithm.An adaptive temperature update function and a dual-threshold are set up to reduce amount of calculation，so can acquire quickly the feature subsets which are more representative.Experimental results show that presented method is effective.