During July 2025, Bhilmes explored enhancements to recurrent neural network workflows in the rwth-i6/i6_models repository by implementing an experimental stateful forward-pass for LSTM components. Using Python and PyTorch, Bhilmes added a forward_with_state method to LstmEncoderV1 and LstmBlockV1, enabling propagation of hidden and cell states across forward passes. After a brief evaluation, the stateful approach was reverted to maintain the stateless behavior required for production stability and compatibility. Bhilmes documented the rationale and outcomes, ensuring repository health and knowledge transfer for future optimization. This work demonstrated thoughtful experimentation and careful consideration of deep learning model requirements.