Piezoresistive soft tactile sensors based on carbon nanotube–polydimethylsiloxane (CNT–PDMS) nanocomposites have emerged as a pivotal technology for flexible electronics and soft robotics owing to their mechanical compliance and high sensitivity. However, their practical deployment remains significantly hindered by nonlinear and rate-dependent hysteresis, which arises from the viscoelastic nature of the elastomeric matrix and the dynamic reconfiguration of internal conductive networks. This path-dependent behavior introduces mapping ambiguity that degrades the reliability of real-time force estimation. Conventional compensation strategies, including phenomenological models such as the Prandtl–Ishlinskii approach, often require complex parameter identification and exhibit limited adaptability to varying dynamic conditions. In this study, a temporal deep learning framework utilizing a stacked long short-term memory network integrated with a sliding-window strategy is presented. Experimental validation on a 0.5 wt% CNT–PDMS sensor under cyclic loading (0.1–2.0 Hz) reveals a substantial reduction in maximum hysteresis error from 12.4 to 1.8% of the full-scale output. The proposed model, optimized via the Adam algorithm, achieves a coefficient of determination of 0.988, ensuring stable performance across diverse dynamic regimes. These findings demonstrate that temporal inference provides a robust and scalable solution for high-precision tactile perception in next-generation soft robotic platforms.