Robots that forget what they just did are getting a fix: a compact memory system called MemBodied.
Researchers built MemBodied, a fixed-size episodic memory for vision-language-action models, the systems that let robots see a scene, understand an instruction, and act. It has two parts: an associative state that tracks interactions across each policy call, and an episode anchor that holds a compact snapshot of the initial scene for reference. Instead of feeding a growing pile of past camera frames back into the model, the usual workaround for tasks that need memory, the robot's action generator conditions on the current input plus these two memory components. On five RMBench tasks that specifically require memory, the approach hit 7.81 times the success rate of a memory-free policy and about triple that of a standard recurrent memory setup.
Manipulation tasks that depend on something the robot saw earlier, like which drawer it already opened, have been a weak spot for these models, and the common fix of stuffing more past frames into context makes inference slower and more expensive. MemBodied beat the strongest memory-augmented baseline in the comparison while using 10 times fewer added parameters, and on the fully observable LIBERO-Long benchmark it improved 5.4 percentage points over the baseline pi_0 policy, reaching 90.6 percent.
The gains are real, but they come from one benchmark suite and one baseline family, so whether this scales past controlled lab manipulation to messier real-world robots is still an open question.