Memory, Energy, and Accuracy Co-Design for Continuous TinyML Sensing: A Critical Narrative Review

Always-on sensing with microcontroller-class hardware creates a design problem in which predictive quality cannot be separated from memory use, power draw, response latency, and sensor availability. This review examines computational co-design for TinyML pipelines that process streaming environmental, acoustic, and industrial signals. It contrasts compact neural architectures, integer quantization, signal-adaptive inference, and device-aware benchmarking through the common question of whether improvements remain useful under continuous operation. A conceptual comparison identifies the costs hidden by accuracy-only evaluations, including peak activation memory, wake-up duty cycle, and calibration drift. The synthesis argues that system-level Pareto evaluation is more informative than comparing isolated neural-network scores. A practical measurement protocol and research agenda are proposed for reproducible assessment under workload shift. No new device measurements or experimental performance results are claimed. The contribution is a critical framing of the evidence needed to substantiate energy-efficient and reliable deployment.