Memory is a hypothesis to test.

In a recurrent design, the state entering a step influences the next computation and is then updated. That explicit continuity can be attractive when a system processes a stream, but it changes the problem: the model must decide what to retain, compress, or discard as new input arrives.

A persistent state is not automatically useful memory. It may discard details needed later, retain distractors, or make optimization harder. A more elaborate update rule can increase capacity while doing little for long-range information use.

Ask which information survives—and when.

A useful experiment isolates the behavior the memory is supposed to provide. Can the model use a dependency introduced many steps earlier? Does it ignore distractors? How does the result change as sequence length grows? Is the effect stable across seeds and task distributions?

Loss curves and parameter counts answer only part of the question. Synthetic recall tasks can isolate specific behaviors; real data tests whether those behaviors matter beyond the toy setting. Baselines and ablations are essential: without comparison, an observed result cannot tell us which component contributed.

Make the update rules inspectable.

For EILER, this means specifying what each state-space block reads and updates, how the memory write is formed, how information is routed, and what the computation costs. Tensor shapes, initialization, numerical stability, gradient flow, and interactions between blocks should be explicit enough to reproduce.

The goal is not to announce a new paradigm. It is to make the hypothesis specific enough that evidence can support it, narrow its scope, or reject it.