To reduce the memory size and fasten the process of searching for a symbol in a Huffman tree, we exploit the property of the encoded symbols and propose a memory-efficient data structure to represent the Huffman tree, which uses memory nd bits, where n is the number of source symbols and d is the depth of the Huffman tree. Based on the proposed data structure, we present an O(logn)-time Huffman decoding algorithm. An adaptive version for single-side growing Huffman tree is also addressed. This version could improve the average performance from logn to \sum\nolimits_{i = 1}^n {\left\lceil {{i \mathord{\left/{\vphantom {i {(h - 1)}}} \right. \kern-\nulldelimiterspace} {(h - 1)}}} \right\rceil } \times {{w_i \log h} \mathord{\left/ {\vphantom {{w_i \log h} {\sum {w_i } }}} \right. \kern-\nulldelimiterspace} {\sum {w_i } }}, where w_i is the frequency for i_th symbol and h is a pre-defined value.