# Divide-and-Remember: Recursive Action-Relevant Memory for Long-Horizon VLA Policies

arXiv: https://arxiv.org/abs/2610.00982 · HTML: https://mlsanity.com/paper/2610.00982/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Reasoning and chain-of-thought in VLA, one of 13 directions in Vision-language-action models.
- Builds on 17 of its 26 references that are in the map.
- First posted on arXiv 1 Oct 2026.

## Placed in

- Vision-language-action models / [Reasoning and chain-of-thought in VLA](https://mlsanity.com/vla/reasoning-and-chain-of-thought-in-vla/)

## Builds on (most cited first, 17 of 17)

- [Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution](https://mlsanity.com/paper/2409.12191/) (2024)
- [π0.5: a Vision-Language-Action Model with Open-World Generalization](https://mlsanity.com/paper/2504.16054/) (2025)
- [End-to-End Training of Deep Visuomotor Policies](https://mlsanity.com/paper/1504.00702/) (2015)
- [DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification](https://mlsanity.com/paper/2106.02034/) (2021)
- [Learning to (Learn at Test Time): RNNs with Expressive Hidden States](https://mlsanity.com/paper/2407.04620/) (2024)
- [Stochastic Beams and Where to Find Them: The Gumbel-Top-k Trick for Sampling Sequences Without Replacement](https://mlsanity.com/paper/1903.06059/) (2019)
- [MemoryVLA: Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic Manipulation](https://mlsanity.com/paper/2508.19236/) (2025)
- [Scalable methods for computing state similarity in deterministic Markov Decision Processes](https://mlsanity.com/paper/1911.09291/) (2019)
- [Contrastive Behavioral Similarity Embeddings for Generalization in Reinforcement Learning](https://mlsanity.com/paper/2101.05265/) (2021)
- [When Do Transformers Shine in RL? Decoupling Memory from Credit Assignment](https://mlsanity.com/paper/2307.03864/) (2023)
- [RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies](https://mlsanity.com/paper/2603.04639/) (2026)
- [HAMLET: Switch your Vision-Language-Action Model into a History-Aware Policy](https://mlsanity.com/paper/2510.00695/) (2025)
- [Conditional Mutual Information for Disentangled Representations in Reinforcement Learning](https://mlsanity.com/paper/2305.14133/) (2023)
- [EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies](https://mlsanity.com/paper/2606.20092/) (2026)
- [Exploiting Causality for Selective Belief Filtering in Dynamic Bayesian Networks (Extended Abstract)](https://mlsanity.com/paper/1907.05850/) (2019)
- [KEMO: Event-Driven Keyframe Memory for Long-Horizon Robot Manipulation with VLA Policies](https://mlsanity.com/paper/2606.23589/) (2026)
- [Recursive Belief Vision Language Action Models](https://mlsanity.com/paper/2602.20659/) (2026)

## Abstract (arXiv)

Vision-language-action (VLA) models struggle on history-dependent manipulation tasks, where the current observation alone does not determine the action, and the policy needs a memory of the history. Existing memory methods decide what to remember by design, for example, keeping frames with large pixel changes, and show inconsistent gains across tasks. We view what to remember as an optimisation problem. From the POMDP formulation of imitation learning, we show that the optimal memory maximises the conditional mutual information I(at; mt mid ot) between the action and the memory given the current observation. Intuitively, this means preserving the action-relevant information in the history that is not already contained in the current observation. Based on our analysis, we propose Divide-and-Remember (D&R), a recursive memory method that learns a memory function mt = M(ht) and scales to long contexts while staying compute-light. It involves two strategies: (1) the selection over the full history is divided recursively into subproblems of top-K selection over 2K tokens, so that fixed-size, lightweight selectors learned end-to-end support an unbounded history; (2) all recursion blocks share one selector, which captures the selection rule common to every block and keeps the method efficient. On RoboMME, a benchmark of 16 long-horizon manipulation tasks that require remembering when, where, what, and how to act, D&R achieves a state-of-the-art average success rate with consistent gains across all four suites under a budget of only 64 tokens; real-robot experiments show the same gain. Code, checkpoints and more results are at https://dnr-memory.github.io/