We watch hours of video daily, yet current AI models are surprisingly bad at remembering what they just saw. As multi-modal systems tackle longer, more complex video streams, a critical flaw has emerged: they lack true memory. This isn't just about forgetting details; it is about how models retain, preserve, and retrieve information when faced with interference. Until now, benchmarks focused heavily on perception and reasoning, completely ignoring the cognitive backbone that allows an AI to actually understand a narrative over time.

Researchers have introduced M$^3$Eval, the first comprehensive framework designed to probe these specific memory dimensions in multi-modal models. Grounded in cognitive psychology, the benchmark uses carefully constructed tasks to isolate how well models retain information, how faithfully they preserve it, and how robust their memory remains when processing new data. By moving beyond simple recognition, this evaluation forces AI systems to demonstrate whether they can truly hold onto distinct pieces of information without them blurring together or fading away.

The results reveal a stark contrast between machine and human memory. Models struggle significantly to maintain separate representations when processing parallel video streams, leading to confused outputs. Surprisingly, their interference patterns differ substantially from human memory, suggesting they do not forget in the same way we do. They are also better at grounding memory in space than in time, and they show limited ability to handle symbolic memory. These findings expose memory as a fundamental but severely underexplored weakness in current architectures.

This benchmark provides a vital resource for the future of AI development. By highlighting these specific failures, it offers clear insights for designing more effective memory mechanisms. If we want multi-modal models to truly understand long-form video, we must stop treating memory as an afterthought and start building systems that can reliably retain and retrieve information just like humans do.