Balancing Memorization and Generalization in RNNs for High Performance Brain-Machine Interfaces