This paper presents a novel complexity-scalable technique to reduce the memory costs of an MPEG-2 decoder by a factor of four to six, by recompressing the I- and P-reference pictures prior to motion-compensated reconstruction. The proposed scheme features some unique characteristics, like robustness for multiple encodings and high memory and memory-bandwidth reduction factors. At the same time, the near-lossless quality obtained is preserved, independently of the ratio between the input and display resolution and transmission/storage bit-rates, because no interference between the memory reduction algorithm and the MPEG-2 decoder occurs. Due to the inherent scalability of the presented algorithm, trade-offs between memory, silicon costs or over-all system costs, and image quality can be easily made. The proposed embedded compression algorithm can also be used for memory and memory-bandwidth reduction in alternative systems, like video coding ICs (e.g. MPEG-4) and memory intensive image processing ICs (e.g. de-interlacing and 100-Hz conversion).