MemBodied: Recurrent Associative Memory for Vision-Language-Action Models
Tej Deep Pala, Navonil Majumder, Bryce Goh, Raphael Yee, Jianfei Yang, Liming Chen, Soujanya Poria
Nanyang Technological University · Griffin Labs · École Centrale de Lyon
Published September 23, 2026 · on the feed September 24, 2026 · ▲ 10 · arXiv 2609.28256
The abstract
Vision-Language-Action models provide a strong foundation for general-purpose robot control, yet a vast majority of policies do not preserve and leverage episode-level information beyond the current observation. This limitation is consequential in history-dependent manipulation tasks that depend on information available only in past observations. Retaining past observations in context can aid in recovering this information, but at the significant cost of ever-growing, bloated context and inference latency. We thus introduce MemBodied, 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 action generation on the current input and the memory components, rather than directly using past observations. Across five evaluated RMBench tasks requiring memory, MemBodied achieves 7.81times the mean success rate of a stateless policy and 2.98times of vanilla recurrent memory, while outperforming the strongest memory-augmented baseline by 1.3times with 10times fewer added parameters. On the fully observable LIBERO-Long suite, it reached 90.6%, a 5.4% improvement over the stateless π_0 policy. These findings support MemBodied as a practical alternative to expanding the policy context for history-dependent manipulation.
The Brief
2:29A short podcast summary of the paper. Written and voiced by AI.
Transcript
MemBodied. Recurrent Associative Memory for Vision-Language-Action Models. By Tej Deep Pala and six others, from Nanyang Technological University, Griffin Labs, and one other institution. Published on the twenty-third of September. Its code repository is worked out: about one hundred and eighty source files, with a readme and a licence.
Vision-Language-Action models provide a strong foundation for general-purpose robot control, yet the vast majority of policies do not preserve and use episode-level information beyond the current observation. That matters in history-dependent manipulation tasks, which depend on information available only in past observations. Keeping past observations in context can recover it, but at significant cost: a context that keeps growing, and rising inference latency. The longer the episode, the more there is to carry and the slower each decision becomes.
MemBodied replaces the growing context with a fixed-size episodic memory. Whatever the episode's length, the memory the policy carries stays the same size. The aim is to keep what the past observations would have told the policy, without keeping the observations themselves.
The memory has two complementary components. An associative state records interactions across policy calls. An episode anchor preserves a compact representation of the initial scene, as a reference. At each policy call, the model conditions its action generation on the current input and these two memory components, rather than directly on past observations. The two parts cover different needs. One tracks what has happened as the episode goes on, and the other keeps a fixed reference to how things looked at the start.
Across five RMBench tasks that require memory, MemBodied reaches seven point eight one times the mean success rate of a stateless policy, and two point nine eight times that of vanilla recurrent memory. It outperforms the strongest memory-augmented baseline by one point three times, with ten times fewer added parameters. On the fully observable LIBERO-Long suite it reaches ninety point six percent, a five point four percent improvement over the stateless pi zero policy. The two settings differ in kind. The RMBench tasks are ones that require memory, while the LIBERO-Long suite is fully observable, so the memory is tested both where history is essential and where the current view already shows everything.
The authors offer MemBodied as a practical alternative to expanding the policy context for history-dependent manipulation. Instead of carrying the whole history, the policy carries two things: a running record of what has happened, and a compact reminder of where the episode began.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Your robot moved the block. Now it has forgotten where the block came from.
Your robot policy looks only at the current camera image and the instruction. It moved a block off its starting pad, and now it must put it back.
The road ahead
- 1 · The problem ~7 min A robot policy that sees only the present view cannot recall facts that are no longer visible, and storing every past frame gets slower and slower.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Replace the growing pile of past frames with a fixed-size memory: a running record of interactions plus a kept copy of the starting scene.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Capture the first scene once, then on each call act from the current view plus memory, and afterwards write what the action caused.
- 6 · The experiment ~6 min You’ll guess: Across five memory-dependent RMBench tasks, how many times the mean success rate of a stateless policy does MemBodied reach?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.