Sensor-based systems must interact with their environments while extracting crucial information, necessary for their performance, from the sensors. In most cases, the projection from the environment to the signal is many-to-one, resulting in irrecoverable information about the environment. To recover this information assumptions must be made. Considering the complexity of the world, we posit that intricate assumptions are necessary for recovering the information. More assumptions require larger knowledge bases, making the performance of the system slower than acceptable. To avoid the crippling effects of large knowledge bases, we accept additional assumptions about the structure of the working environments and the interaction of systems with their environments along different dimensions. These assumptions allow systems to dynamically hide large portions of knowledge that are irrelevant at a given time. We call this approach knowledge caching. We introduce an implementation of this approach in the context-based caching (CbC) technique in which knowledge items are swapped based on precompiled relations between knowledge items. This technique enhances system performance providing it with the right information at the right time.