kei-neuromem — a memory that grows.

Instead of one large model trying to represent language at every scale at once, kei-neuromem splits the job: a committee of small, per-level nets segments a raw character stream into a growing, typed multi-order token graph — char, syllable, morpheme, word, phrase — storing a connection only when it is confident about it. The memory is a module that feeds a downstream processor, not a replacement for one.

This is deliberately not the lab's earlier, measured-refuted idea of a matrix-valued token substrate competing with standard embeddings — that comparison showed no benefit at matched parameter counts. kei-neuromem is a memory and an honest-segmentation front end, not a token substrate trying to out-perform standard tokenization end-to-end.

The store is built on a modern-Hopfield energy function — a landscape with many stable minima, one per stored pattern, instead of a single global attractor. That is a mathematical stability property, not a tuning trick: a recalled pattern settles into the nearest minimum and stays there, so the memory can hold many distinct items at once by construction, rather than by hoping a heuristic happens to behave.

An initial validation probe measured that store successfully holding multiple items without collapsing into a single attractor, while a control condition (the lab's earlier substrate) reproduced the collapse it is designed to avoid. One further falsifier tested the wrong control mechanism and needs a re-run against the correct one. Early-stage — reported here as measured, not as a finished system.