Smart environments equipped with heterogeneous sensors—wearables, cameras, microphones, and ambient devices—generate continuous streams of multimodal data that can enable intelligent monitoring of human health, behavior, and well-being. However, current multimodal monitoring systems operate with a fixed sensing strategy: all available sensors capture data at their maximum capacity regardless of the user's context, activity, or the system's actual information needs. This alwayson approach leads to excessive energy consumption that drains wearable batteries within hours, overwhelming data volumes that strain storage and communication bandwidth, redundant information capture across overlapping modalities, and unnecessary privacy intrusion during contexts where highfidelity sensing is not needed. These limitations represent a fundamental barrier to the practical, longterm deployment of multimodal monitoring systems in real-world settings. This project proposes AdaptSense, an adaptive multimodal sensing framework that intelligently determines what to sense, when to sense, and at what fidelity, based on the user's current context and the system's monitoring objectives. The framework operates through three interconnected components. First, a context recognition module continuously infers the user's current state—activity, physiological condition, and environmental setting—from lightweight, low-power sensor inputs. Second, a sensing policy network, trained using reinforcement learning, dynamically selects which sensors to activate, at what sampling rates, and for how long, optimizing the trade-off between monitoring accuracy and resource cost (energy, bandwidth, storage, privacy). Third, a multimodal inference engine processes the adaptively sampled data to perform the target monitoring tasks, operating robustly under the variable and incomplete sensor configurations that result from the adaptive sampling decisions. The framework is applicable across multiple domains where sustained multimodal monitoring is needed. In health monitoring, AdaptSense enables long-term remote patient monitoring by activating highfidelity multimodal capture (physiological signals, speech, vision) only when the context warrants it— such as detecting anomalous physiological patterns or behavioral changes—while defaulting to minimal low-power sensing during routine, stable periods. In smart home and ambient assisted living, the framework adapts sensing to the resident's daily routine, activating activity recognition and safety monitoring during high-risk periods while reducing sensing during rest or absence, supporting elderly aging in place with minimal intrusiveness. The framework will be validated on at least one of these application domains using established public datasets, with extensibility to additional domains demonstrated through cross-domain evaluation.