We present a new algorithm for training neural networks with binary activations and multi-level weights, which enables efficient processing-in-memory circuits with eNVM. Binary activations obviate costly DACs and ADCs. Multi-level weights leverage multi-level eNVM cells. Compared with previous quantization algorithms, our method not only works for feed-forward networks including fully-connected and convolutional, but also achieves higher accuracy and noise resilience for recurrent networks. In particular, we present a RNN trigger-word detection PIM accelerator, whose modeling results demonstrate high performance using our new training algorithm.