Non-binary (NB) low-density parity check (LDPC) codes offer stronger error correcting capability than binary LDPC codes. However, the improvement is accompanied by a considerable increase in the decoding complexity and a huge memory requirement, especially in the check node (CN). To overcome these limitations, we propose a memory-reduced NB-LDPC decoding method with an accumulative bubble check (a-bubble check) algorithm. The proposed a-bubble check reduces the memory requirements of the CNs by using the differences between the log-likelihood ratio entries. In addition, we propose a non-uniform memory structure to further decrease the memory capacity of the CN processor. For the GF(64)(160, 80) NB-LDPC code, the proposed memory-reduced NB-LDPC decoding architecture achieves a memory requirement reduction of 32.4% for the CNs and 16.2% for the entire decoder without degrading the error correction capability.