We present significant evidence of halo assembly bias for SDSS redMaPPer galaxy clusters in the redshift range $[0.1, 0.33]$. By dividing the 8,648 clusters into two subsamples based on the average member galaxy separation from the cluster center, we first show that the two subsamples have very similar halo mass of $M_{rm 200m}simeq 1.9times 10^{14}~h^{-1}M_odot$ based on the weak lensing signals at small radii $R<sim 10~h^{-1}{rm Mpc}$. However, their halo bias inferred from both the large-scale weak lensing and the projected auto-correlation functions differs by a factor of $sim$1.5, which is a signature of assembly bias. The same bias hypothesis for the two subsamples is excluded at 2.5$sigma$ in the weak lensing and 4.4$sigma$ in the auto-correlation data, respectively. This result could bring a significant impact on both galaxy evolution and precision cosmology.