The channel polarization under the channel noise with memory is comprehensively studied. With the help of the genie-aided channel, we prove that the polarized channels also converge to extremal channels under the standard polar codes structure. More importantly, the ratio of the perfect channel can be larger than $I(W)$ which is the capacity of the original channel. However, the polarization rate is shown to be slower than the binary-input discrete memoryless channel (DMC) case. Specifically, the upper bound of the block error is $\mathcal{O}(L^{-c_0})$ where $L$ is the block length and $c_0$ is some positive constant. Furthermore, the upper and lower bound of the gap between the capacity and cutoff rate is investigated when the block length is finite, which is more useful for practical applications.