VideoMemBench: 20+ MLLMs Tested on Cross-Turn Video Memory — Top Model Caps at 45.7%

Hour-long multi-turn video dialogue is becoming a primary MLLM use case, yet no benchmark previously tested whether models can recall and build on what they said or saw in earlier turns of the same session. A 45.7% ceiling across 20+ models—including frontiers—is a concrete and reportable gap.