—Standard Recurrent Neural Networks (RNNs) compress historical context into opaque hidden state vectors, while Transformers rely on additive positional biases and O(N2) attention. In this paper, we introduce a lightweight memory architecture that stores token in stances explicitly and retrieves them via Direct-Jump routing. Instead of standard additive attention, we utilize Multiplicative Structural Gating, where inverse hop distance acts as a strict multiplicative prior on memory retrieval.Through ablation studies on a sequence memorization task,we demonstrate that this combination successfully learns sequence representations with 100% accuracy, whereas standard additive attention fails. Furthermore, we in vestigate compressing this instance memory into a fixed V ×V token-type graph. We show that this compression fails because a fixed token-type matrix cannot preserve the sequence order of repeated token instances, and that augmenting it with a predictive surprise gate causes catastrophic runaway feedback loops.
Hasmukh Madanlal Sutar (Fri,) studied this question.