Cornell, UC Berkeley, Harvard and Impossible Inc. · Long Video Memory

MosaiChunk

A memory method for autoregressive video generation past a frozen generator's context window. Instead of retrieving whole past chunks, it splits each chunk's key-value cache into sections with descriptors held in CPU memory; the newest chunk queries those descriptors and assembles the best-scoring sections into a mosaic within a fixed memory budget. Only the descriptor encoder is trained, by self-distillation with no ground-truth video. Tested on MiniMax-H3 and LingBot-World-Infinity with a new RememBench benchmark, it reports 0.899 CLIP with one chunk of memory against 0.841 for two-chunk whole-chunk retrieval, self-reported. Paper arXiv 2610.02153 (1 October), code 8 October.