Developed an adaptive frequency-based cache admission feature for the kvcache-ai/Mooncake repository, focusing on optimizing local hot cache usage under high concurrency. The solution introduced frequency-based promotion of frequently accessed keys, ensuring that hot data remains readily available while reducing cache contention. By refining LRU lock handling, the implementation improved performance and scalability in multi-threaded environments. The work leveraged C++ for efficient algorithm design and incorporated advanced data structure techniques to manage cache state. This targeted approach addressed the challenge of balancing cache hit rates with lock contention, resulting in a more responsive and robust caching layer for demanding workloads.