Vision-Language-Action (VLA) models have become a strong foundation for general-purpose robot control, but most policies discard episode-level information once the current observation has passed. That gap matters in history-dependent manipulation tasks, where the cues needed to act are only available in earlier observations.
Keeping past observations in context can recover that information, yet it comes at a steep price: an ever-growing context window and higher inference latency. The authors of MemBodied propose a fixed-size episodic memory as an alternative.
The memory has two complementary parts: an associative state that records interactions across policy calls, and an episode anchor that keeps a compact representation of the initial scene as a reference. At each policy call, the model conditions action generation on the current input plus these memory components, rather than directly consuming past observations.
Across five evaluated RMBench tasks that require memory, MemBodied reports 7.81× the mean success rate of a stateless policy and 2.98× that of vanilla recurrent memory. It also outperforms the strongest memory-augmented baseline by 1.3× while adding 10× fewer parameters. On the fully observable LIBERO-Long suite, it reached 90.6%, a 5.4% improvement over the stateless π_0 policy.
The authors position MemBodied as a practical alternative to expanding policy context for history-dependent manipulation.