Convergent Architectures: Computational Vossels and Compute-in-Memory State Space Models as a Unified Framework for Edge AI

Abstract

Two independent lines of research, developed without knowledge of each other, have converged on the same fundamental problem from opposite directions. Our Computational Vossel architecture (Passmore, 2025, 2026) provides a biomimetic software framework in which each spatial data element simultaneously stores and processes sensor measurements, organized by seven agent classes modeled on the organelles of the eukaryotic cell. Zhang et al. (2026) provide a co-designed hardware system in which state space models are physically mapped onto resistive RAM crossbar arrays, with tungsten oxide memristors implementing state decay through their native conductance physics rather than through digital computation. Neither work, by itself, is complete. The vossel architecture needs a physical substrate that matches its storage-compute fusion principle; the CIM-SSM hardware needs an organizing computational framework that specifies what to compute, how to route results, and how to manage multi-modal sensor fusion across spatial elements. Each provides exactly what the other lacks. This paper makes the convergence thesis precise. I show that the seven vossel agent classes map directly and nontrivially onto the primitive operations of the CIM-SSM hardware: the Endoplasmic Reticulum agent implements the SSM state evolution equation; the Golgi agent implements the output projection matrix C; the Nucleus agent configures decay rates and active channels; the Mitochondria agent maps to independently power-gated CIM cores; the Receptor agent implements selective input weighting through the input projection matrix Bˉ ; and the mipvol hierarchy maps to cascaded SSM blocks. The correspondence is not analogical but structural: the mathematical formalism of each vossel agent maps to a specific matrix operation or physical process in the CIM-SSM chip. Together, the two architectures constitute a complete, biologically-grounded edge AI stack that eliminates the von Neumann bottleneck for real-time sensor processing. Energy analysis shows the combined system achieves orders-ofmagnitude efficiency advantages over GPU-based edge inference for asynchronous event-stream workloads, with direct applicability to tactical ISR platforms operating under severe size, weight, and power constraints.