KeyRec Framework for Bounded Visual Memory in Video Models
September 25, 2026
KeyRec is a training-free framework designed to reduce the computational cost of long-video VLM inference. It uses a dual-structure approach: a visual cache for fine-grained recent observations and a structured event bank for historical evidence.
HOW THIS AFFECTS YOU
●
builderYou can reduce inference costs for long-video processing without retraining your vision-language model.
●
researcherThis provides a training-free method to manage token accumulation in continuous video streams.