A paper posted on arXiv introduces the design-model framework, a method for deriving efficient recurrent sequence maps from explicit assumptions about memory. The framework is described as a way to turn such assumptions into concrete computational structures for sequence modeling.

The core mechanism writes evidence into memory using exact Bayesian filtering, meaning the model updates its memory state according to Bayesian principles as new data arrives. The abstract also mentions a query component, though the full details of that part are not included in the available text.

By grounding sequence models in explicit memory assumptions, the framework offers a principled alternative to purely heuristic recurrent designs. The paper is available on arXiv under identifier 2605.31163.