As the number of cores continuously grows, processor designers are considering non coherent memories as more scalable and energy efficient alternatives to the current coherent ones. The Java Memory Model (JMM) requires that all cores can access the Java heap. It guarantees sequential consistency for data-race-free programs and no out-of-thin-air values for non data-race-free programs. To implement the Java Memory Model over non-cache-coherent and distributed architectures Java Virtual Machines (JVMs) are most likely to employ software caching. In this work, i) we provide a formalization of the Java Memory Model for non-cache-coherent and distributed memory architectures, ii) prove the adherence of our model with the Java Memory Model and iii) evaluate, regarding its compliance to the Java Memory Model, a state-of-the-art Java Virtual Machine implementation on a non-cache-coherent architecture.