MemBodied Beats Stateless VLA Policy 7.81x on RMBench Tasks

On five memory-heavy RMBench tasks, MemBodied reaches 7.81 times the mean success rate of a stateless policy, according to a preprint posted September 23 by Tej Deep Pala, Navonil Majumder, Bryce Goh, Raphael Yee, Jianfei Yang, Liming Chen and Soujanya Poria.

The pitch is a swap: instead of stuffing more past observations into the policy's context, keep a small, fixed slot of episode state. The paper introduces "a fixed-size episodic memory with two complementary components: an associative state that records interactions across policy calls and an episode anchor that preserves a compact representation of the initial scene as a reference." At each policy call, the model conditions on the current input and those memory components rather than on raw history.

Against a vanilla recurrent memory the multiplier is 2.98x. Against what the authors call "the strongest memory-augmented baseline" MemBodied claims a 1.3x edge "with 10x fewer added parameters." On the fully observable LIBERO-Long suite it reaches 90.6%, which the abstract frames as "a 5.4% improvement over the stateless pi_0 policy."

The motivation the paper gives for the whole design is blunt: retaining past observations in context helps, but at "the significant cost of ever-growing, bloated context and inference latency." The abstract does not publish wall-clock latency figures alongside the accuracy gains, and does not name the specific memory-augmented baselines it outperforms.

It lands amid a steady drip of episode-memory-for-robots work our tracker has been logging in the robotics feed, including a same-day preprint on generating diverse VLA training data from frontier coding agents.