With the widespread use of sensors in smart devices and robots, there is a growing expectation for edge computing that processes data not on distant cloud servers but also on or near interactive devices to store their data with low latency access. To satisfy these requirements, we consider a new edge computing system that consists of a hybrid main memory with a KVS (Key-Value-Store) server utilizing the DRAM and nonvolatile main memory (NVM). It provides large-capacity cache memory in a server, supporting high-speed processing and quick response for sensor nodes. However, since existing KVS servers are not designed for NVM, there are less satisfactory implementations that achieve low response time and high throughputs. We propose a novel hybrid KVS server that is designed and implemented on the Memcached distributed memory-caching system, which dynamically moves cached data between two types of memory devices according to access frequency in order to achieve a low latency compared to the existent approaches. We developed a Dual-LRU (Least Recently Used) structure for it. Evaluation was performed using a real machine equipped with NVM. The result showed the proposed method successfully reduced the response time and improves access throughputs.