mlsanityData to 14 Oct 2026

AURA: Action-Gated Memory for Robot Policies at Constant VRAM

arXiv 2606.02775 ↗First posted 1 Jun 2026cs.AI cs.AR cs.DCcited by 2

On the map: Efficient LLM inference and serving → Visual token pruning in multimodal models

Abstract (arXiv)

The KV-cache is the right memory for datacenters but the wrong memory for robots. Datacenter inference batches many short requests and resets them, amortizing an attention cache across a crowd. Embodied agents instead run one long, non-resetting episode on bandwidth-limited edge hardware, where high-bandwidth memory and flash are scarce, flash has finite write endurance, and memory writes rather than compute can become the binding constraint. AURA-Mem (Action-Utility Recurrent Adaptive Memory) targets this regime. It wraps a frozen vision-language-action backbone with a constant-size recurrent memory and a learned gate that writes only when the current observation would change the next action: memory that knows when to stay silent. Unlike reconstruction-based memory, the gate is trained directly against a closed-loop action-error signal. Its inference state is fixed at 4,224 bytes regardless of horizon, while a KV-cache grows to 6,061 times larger at 100,000 steps. On a controlled synthetic benchmark, AURA-Mem matches the best O(1) baseline in accuracy while using 5.19-6.13 times fewer writes, and up to 9.19 times fewer writes on easier configurations. Budget-matched random and periodic schedules do not recover this gain, isolating the benefit to the action-surprise signal. On a trained closed-loop OpenVLA-OFT 7B panel on LIBERO-Long (n=60 episodes per arm), the gate does not hurt success: AURA-Mem matches the ungated base policy (0.233) and slightly exceeds an always-write KV arm (0.217), while using 7.0 times fewer writes and constant memory. We also instantiate an approximate-information-state value-loss bound as a methodology demonstration; at this scale, the bound is vacuous rather than a guarantee.

Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Placed in Visual token pruning in multimodal models, one of 12 directions in Efficient LLM inference and serving. Builds on 56 of its 67 references that are in the map. First posted on arXiv 1 Jun 2026.

Builds on

56 of 67 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models11 Jun
  2. 2024VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration29 Oct

Other references · 54

  1. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,944
  2. 2023Mamba: Linear-Time Sequence Modeling with Selective State Spacescited by 4,718
  3. 2022FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awarenesscited by 3,673
  4. 2021Efficiently Modeling Long Sequences with Structured State Spacescited by 2,465
  5. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,186
  6. 2020Rethinking Attention with Performerscited by 1,601
  7. 2021Decision Transformer: Reinforcement Learning via Sequence Modelingcited by 1,546
  8. 2024Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Dualitycited by 1,336
  9. 2017Curiosity-driven Exploration by Self-supervised Predictioncited by 1,194
  10. 2020Mastering Atari with Discrete World Modelscited by 1,007
  11. 2018Exploration by Random Network Distillationcited by 993
  12. 2016Deep Variational Information Bottleneckcited by 930
38 more
  1. 2024SnapKV: LLM Knows What You are Looking for Before Generationcited by 880
  2. 2023H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Modelscited by 832
  3. 2023RWKV: Reinventing RNNs for the Transformer Eracited by 827
  4. 2022Simplified State Space Layers for Sequence Modelingcited by 788
  5. 2024KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cachecited by 637
  6. 2023Retentive Network: A Successor to Transformer for Large Language Modelscited by 613
  7. 2024KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantizationcited by 591
  8. 2023Gated Linear Attention Transformers with Hardware-Efficient Trainingcited by 514
  9. 2021RMA: Rapid Motor Adaptation for Legged Robotscited by 508
  10. 2016RL2: Fast Reinforcement Learning via Slow Reinforcement Learningcited by 503
  11. 2021Linear Transformers Are Secretly Fast Weight Programmerscited by 502
  12. 2020Hopfield Networks is All You Needcited by 475
  13. 2023Model Tells You What to Discard: Adaptive KV Cache Compression for LLMscited by 474
  14. 2023Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Timecited by 439
  15. 2020Learning Invariant Representations for Reinforcement Learning without Reconstructioncited by 436
  16. 2024Learning to (Learn at Test Time): RNNs with Expressive Hidden Statescited by 355
  17. 2024Titans: Learning to Memorize at Test Timecited by 328
  18. 2024Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inferencecited by 235
  19. 2024Simple linear attention language models balance the recall-throughput tradeoffcited by 208
  20. 2019DeepMDP: Learning Continuous Latent Space Models for Representation Learningcited by 207
  21. 2016Using Fast Weights to Attend to the Recent Pastcited by 170
  22. 2025Test-Time Training Done Rightcited by 130
  23. 2020Data Movement Is All You Need: A Case Study on Optimizing Transformerscited by 108
  24. 2024RoboMamba: Efficient Vision-Language-Action Model for Robotic Reasoning and Manipulationcited by 103
  25. 2024AI and Memory Wallcited by 92
  26. 2025VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Cachingcited by 86
  27. 2020Approximate information state for approximate planning and reinforcement learning in partially observed systemscited by 70
  28. 2025MemER: Scaling Up Memory for Robot Control via Experience Retrievalcited by 67
  29. 2025ATLAS: Learning to Optimally Memorize the Context at Test Timecited by 66
  30. 2025It's All Connected: A Journey Through Test-Time Memorization, Attentional Bias, Retention, and Online Optimizationcited by 64
  31. 2023POPGym: Benchmarking Partially Observable Reinforcement Learningcited by 62
  32. 2026MEM: Multi-Scale Embodied Memory for Vision Language Action Modelscited by 61
  33. 2026RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policiescited by 49
  34. 2026Gated DeltaNet-2: Decoupling Erase and Write in Linear Attentioncited by 33
  35. 2026Learning to Evict from Key-Value Cachecited by 14
  36. 2026Gated Memory Policy: In-Context Memorization and Adaptationcited by 14
  37. 2025Memo: Training Memory-Efficient Embodied Agents with Reinforcement Learningcited by 9
  38. 2025ELMUR: External Layer Memory with Update/Rewrite for Long-Horizon RL Problemscited by 6

50 most cited of 54 All 54 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026AEGIS: A Backup Reflex for Physical AIcited by 3
  2. 2026Memory as a Wasting Asset: Pricing Flash Endurance for Embodied Agents, and the Limits of Doing Socited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026rMuscle: Robotic Muscle Memory for Efficient Vision-Language-Action Model Inference
  2. 2026AGM: Achievement-Grounded Memory for Closed-Loop Agents with Frozen VLA Policies
  3. 2026T2Mem: Learning Test-Time Memory for Robotics ↗
  4. 2026StateMem: Single-State Residual Memory with Adaptive Inference for Vision-Language-Action Policies
  5. 2026ECoMEM: Explicit Concept Memory for Memory-Dependent Robot Control ↗
  6. 2026Agent Memory Below the Prompt: Persistent Q4 KV Cache for Multi-Agent LLM Inference on Edge Devices
  7. 2025KV-Efficient VLA: A Method to Speed up Vision Language Models with RNN-Gated Chunked KV Cache
  8. 2026Beyond LLM Serving: Characterizing Vision-Language-Action Workloads for Embodied AI System Design