2(types of explicit self-esteem) ×2(types of implicit self-esteem) ×2(levels of encoding) ×3(valence of words) mixed design was mainly adopted in this study to explore memory bias of valences of the words on two levels of encoding of 89 participants with different self-esteem structures. The results were as follows: (1)In shallow level of encoding group, participants with low implicit self-esteem had more positive memory bias than those with high implicit self-esteem. But there was no difference between participants with high and low explicit self-esteem.(2)In deep level of encoding group, different self-esteem structures had different influences on unconscious retrieval of positive words.(3)All participants had positive bias.