1R memristor-based crossbars provide a compact and energy-efficient platform for in-memory computing but suffer from sneak currents, which are typically viewed as a reliability issue. In this work, we reinterpret sneak currents as a potentially useful computational phenomenon and leverage their spatiotemporal dynamics to construct physical reservoirs (a type of recurrent neural networks). We propose an evolutionary synthesis framework that co-optimizes memristor states and input connections to control sneak current flow, enabling adaptive input masking and modular circuit structures. Experimental results on time-series prediction benchmarks show that the evolved memristive reservoirs, which deliberately exploit sneak currents as additional dynamical states, outperform existing software- and hardware-based models in prediction accuracy while maintaining reliable computation.