VLA memory debate: text-level long history plus sliding-window frames for real-time control

YouJiacheng · x · 2026-11-23

In a discussion on VLA memory architecture, YouJiacheng argues humans clearly don't remember every frame, so a robot's long-term history is likely sufficient at the concept/text level. He proposes VLAs maintain a sliding window of recent frames for real-time perception and control, directed by an LLM that holds long but non-visual history.

Related event: VLA Memory Design: LLM for Long-Term History, Sliding Window for Control(3 posts)→

Original post →

More from Embodied

Embodied channel →