The amount of memory resources in high-performance computing systems (HPC) has increased in the past decade to cope with evolving and diverse workloads. Meanwhile, large-scale studies on production clusters indicate that memory utilization is often low, meaning that money and energy are wasted. A disaggregated memory system can provide HPC applications with on-demand memory capacity and bandwidth, in contrast to today's monolithic designs, where compute and memory are tightly coupled. To understand the opportunities and challenges for HPC applications, we developed an emulator to evaluate HPC applications on disaggregated memory. We used the emulator to prototype a configurable and modular memory system to study the performance of seven scientific applications in various configurations. 5/7 evaluated applications showed minor to moderate performance impact even with up to 75% memory footprint on disaggregated memory. Bandwidth-sensitive applications show large performance degradation, but adding remote links mitigated the performance impact. However, interference among co-running jobs could become a major challenge to overcome for adoption in production systems.