This paper presents a novel, ultra-low-cost medicine reminder system developed on the ESP32-S3 microcontroller, designed to support medication adherence through multimodal alerts and intelligent user interaction. Reminders are issued via an OLED display, buzzer, and pre-recorded voice prompts, while adherence is verified using physical sensors. The key innovation lies in the integration of a fully on-device, cascading TinyML pipeline: a Keyword Spotting (KWS) model triggers a Spoken Language Understanding (SLU) network for intent classification, enabling natural, context-aware voice interaction. Both models are optimized for the ESP32-S3, ensuring low-power, real-time inference without cloud dependency. The system further incorporates multimodal compliance verification (voice + lid + hand sensing) and caregiver notifications, achieving 96.4% KWS and 94.1% SLU accuracy with sub-100 ms latency and minimal memory footprint. Compared to existing low-cost solutions, our design provides higher adherence accuracy, improved robustness in noisy or resource-constrained environments, and enhanced accessibility for elderly or sensory-impaired users. Its modular, scalable architecture enables offline operation, privacy preservation, and adaptive functionality, offering a uniquely practical and intelligent healthcare companion.