Task-Focused Memorization for Multimodal Agents

Researchers introduced TaskMem, a two-phase memory framework for multimodal agents that learns which observations are worth storing long term.

It first learns general memory quality, then adapts a memorization policy to recent tasks using a reward model.

The team converted VideoMME, EgoLife, and EgoTempo into streaming benchmarks to test task-relevant memory in realistic settings.

Built on Qwen3-VL-30B-A3B, TaskMem improved VQA accuracy and addressed a key bottleneck in continual multimodal learning.