Question
Can a small feed-forward network use recent input history through an explicit accumulator?
Approach
The model adds an accumulator buffer to a PyTorch multilayer perceptron. Accumulated signals decay as new inputs arrive, allowing the current computation to depend on earlier activity.
Training and stream-handling code explore the model on sequential inputs.
Open question
The decay rule sets the timescale of the retained signal. Its usefulness needs comparison with simpler stateful baselines under the same data order and state-reset rules. The current project does not establish an advantage over those alternatives.