In resource-constrained edge devices, conventional general-purpose memory allocation mechanisms are prone to high fragmentation rates and increased peak memory usage, leading to low resource utilization. To address this issue, this research proposes a large memory allocator based on program behavior characteristics. The allocator employs a 2-level light-weight shared cache architecture, where radix trees and bitmaps are utilized for efficient free block searching and coalescing. The lightweight cache design is capable of reducing peak memory consumption and improving memory management efficiency. To accommodate diverse program behavior characteristics, both static program features and dynamic execution data are collected. Key memory usage patterns are extracted from this data to drive a genetic algorithm for the automatic search of optimal configuration parameters. These parameters are then applied during program execution to guide tiered reclamation and memory pre-allocation, thereby reducing memory fragmentation and allocation latency. Experimental results indicate that, in scenarios with frequent large memory allocations, the proposed allocator, even with default configurations, achieves lower peak memory usage than most general-purpose allocators. After adaptive tuning, the fragmentation rate can be further reduced to below 5%, with a negligible increase in allocation delay of only 2.3%. Benefiting from a fine-grained locking mechanism, the proposed allocator maintains good scalability even under high concurrent multi-threaded workloads. This work provides a novel approach for achieving high-performance, low-fragmentation memory management in resource-constrained scenarios.