The document proposes a Discreteness-Aware Approximate Message Passing (DAMP) algorithm for reconstructing discrete-valued vectors from underdetermined linear measurements. DAMP extends existing AMP algorithms to handle discrete variables by incorporating probability distributions of the elements. The algorithm is analyzed using state evolution to derive conditions for perfect reconstruction. A Bayes optimal version of DAMP is also developed by minimizing mean squared error. Simulation results demonstrate improved reconstruction performance compared to conventional methods.