Physical Intelligence adds dual-memory stack to VLAs

This is a meaningful step from flashy robot demos toward systems that can actually track multi-stage work over time.

• –MEM splits memory into two channels: dense recent observations and compressed textual task history.
• –The approach is designed to reduce token load, which matters for real-time robotic control loops.
• –The blog and paper frame gains around long-horizon kitchen-style tasks and in-context adaptation after failed attempts.
• –If these results transfer broadly, VLA progress may hinge more on memory design than on scaling raw policy size.