Large switching current and long switching time have significantly limited the adoption of spin-transfer torque random access memory (STT-RAM). Technology scaling, moreover, makes it very challenging to reduce the switching current while maintaining the reliability of magnetic tunneling junction (MTJ) to be similar to that of the earlier generations. In this work, we shall exploit a key insight that in the most on-chip caches where STT-RAM is most likely to be deployed, the lifespan of the data stored in the memory cells is much shorter than the data retention time requirement assumed in STT-RAM development, namely, 4~10 years. We also quantitatively investigated the possibility of trading off MTJ nonvolatility for improved switching performance, e.g., the switching time and/or current, under architectural level guidance. We further proposed and evaluated a hybrid memory design technique that partitions the on-chip STT-RAM cache into two parts with different nonvolatility performances so as to better fit the diverse retention time requirements of different data sets.