A research framework named Homer can now reason across hour-long videos without losing the plot.
Most multimodal AI models handle short clips well but fall apart when asked to answer questions about footage that runs for an hour or more. The core problem is memory: existing online systems either compress visual detail too aggressively or organize what they retain by time alone, ignoring cause and effect. Homer, short for Hierarchical Online Memory Exploration and Reasoning, tackles this by building memory in layers — raw frames at the bottom, recurring entities in the middle, and causally linked events at the top. An agentic reasoner then navigates those layers the way a person skims notes, locating the right scene, pulling specific details, and checking its own work before answering.
The benchmark numbers are meaningful. Homer beat the previous best agent method by 5.5 points on M3-Bench-robot, 10.8 points on M3-Bench-web, and 4.4 points on Video-MME-Long. Importantly, the gains held across three different underlying language models, which suggests the architecture is doing real work rather than being tuned to one backbone's quirks. That model-agnostic quality is what makes this more than a leaderboard footnote.
Long-video understanding is becoming a genuine competitive frontier — Google, OpenAI, and several startups have all claimed progress here in the past year. Homer's causal memory structure is a credible engineering answer to a problem those systems have mostly papered over with longer context windows.