Cache is used to improve performance in object-relational mapping. Data consistency maintenance between cache and server is the key factor of system reliability. In this paper, a granularity adaptive consistency maintenance scheme that is initiated by middle layer is proposed. This scheme considers several data attributes including storage granularity, updating frequency and size. After combining the number of cache items, it uses TTL and on-demand request respectively. Experimental results show that the proposed policy can effectively reduce network overhead and load. Data efficiency can also be guaranteed.