pith. sign in

arxiv: 2606.02775 · v1 · pith:BKGEN2U4new · submitted 2026-06-01 · 💻 cs.AI · cs.AR· cs.DC· cs.PF· cs.RO

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

Josef Chen This is my paper

Pith reviewed 2026-06-28 14:28 UTC · model grok-4.3

classification 💻 cs.AI cs.ARcs.DCcs.PFcs.RO
keywords action-gated memoryrobot policiesconstant memoryKV-cachevision-language-actionLIBERO benchmarkrecurrent memoryedge deployment
0
0 comments X

The pith

A learned action gate lets robot policies use constant memory by writing only when observations change the next action.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

AURA-Mem wraps frozen vision-language-action models with a fixed-size recurrent memory controlled by a gate trained on closed-loop action errors. The gate writes only when the current observation would alter the subsequent action, keeping memory at 4,224 bytes no matter how long the episode runs. This matches the success rate of an ungated base policy on LIBERO-Long while cutting writes by a factor of seven and avoids the growing KV-cache that reaches thousands of times larger size.

Core claim

AURA-Mem is a constant-size recurrent memory with an action-utility gate that updates only on observations predicted to change the next action, trained directly against closed-loop action-error rather than reconstruction. On a synthetic benchmark it matches the accuracy of the best constant-memory baselines with five to nine times fewer writes. On the LIBERO-Long benchmark with a 7B OpenVLA-OFT policy it preserves the base success rate of 0.233 while using seven times fewer writes and fixed memory footprint.

What carries the argument

The action-surprise gate trained on closed-loop action-error that decides writes to a fixed 4,224-byte recurrent memory.

If this is right

  • Memory footprint remains fixed at 4,224 bytes for any episode length up to at least 100,000 steps.
  • Success rate on LIBERO-Long stays at 0.233, matching the ungated policy and exceeding the always-write KV version at 0.217.
  • Synthetic benchmark accuracy matches the best O(1) baseline while requiring 5.19 to 9.19 times fewer writes.
  • Budget-matched random and periodic write schedules fail to match the performance of the action-surprise gate.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The gate may generalize to other long-running embodied agents where flash endurance limits total writes.
  • Training the gate on action error rather than reconstruction could apply to other memory compression problems in sequential decision making.
  • Constant memory enables deployment on hardware with strict VRAM constraints that growing caches cannot meet.

Load-bearing premise

That training the gate on closed-loop action-error signals will maintain policy performance when applied to new tasks and real robot episodes.

What would settle it

Running the AURA-Mem policy on a longer or different robot task where its success rate falls below the ungated baseline while memory stays constant.

Figures

Figures reproduced from arXiv: 2606.02775 by Josef Chen.

Figure 1
Figure 1. Figure 1: AURA-Mem in one picture: memory that knows when to shut up. The whole system is a learned write gate plus a bounded fast-weight state wrapped around a frozen VLA back￾bone. At each tick the observation ot is summarised by the frozen backbone (OpenVLA-OFT 7B, weights unchanged) into a latent token zt=proj(ht). An action-error write gate gt=1[pt>τ ] decides whether to write: it fires only on surprise, so mos… view at source ↗
Figure 2
Figure 2. Figure 2: AURA-Mem single control-step datapath. At each tick t the frozen backbone summarises the raw observation into a latent token zt ∈ R d . Three linear projections emit a query q=θQz, key k=θKz, and value v=θV z; the fast-weight state W ∈ R dk×dv is read first via o=q ⊤W to produce the memory output before any write occurs. The surprise scalar s=∥kW − v∥ 2 2 (the inner TTT reconstruction error, detached from … view at source ↗
Figure 3
Figure 3. Figure 3: O(1) constant-shape state vs. growing KV-cache across a long episode. Top: AURA-Mem unrolled over T steps: the fast-weight tensor W maintains a fixed shape [dk × dv] at every step; only the contents of W evolve, and only on gate-selected steps (sparse ticks, filled circles); the resident inference-state footprint is 4,224 bytes throughout (formula (dkdv + dv)×batch×4, dk=dv=32), confirmed constant across 1… view at source ↗
Figure 4
Figure 4. Figure 4: Write-bandwidth vs. accuracy frontier (noisy long recall, T=96, N=64, 4,000 training steps; Wong colorblind-safe palette; error bars: 95% t-interval). Each point is one variant’s mean task success plotted against its mean write bandwidth (writes/sec, log scale) at the highest evaluated state budget (N=64); lower bandwidth is preferable for DRAM/HBM wear and energy cost. AURA-Mem (blue star) achieves 9.19× … view at source ↗
Figure 5
Figure 5. Figure 5: Task success rate vs. state budget N on the hard noisy long recall con￾figuration (nkeys=16, nvals=8, nbindings=16, distractor= 0.5, overwrite= 0.4, T=128; up to 6 seeds per cell; chance floor 0.125; shaded bands: 95% t-CI). AURA-Mem (solid blue) matches fixed size state (solid green) at every tested budget (N∈{8, 16, 24, 32}): accuracy gaps are ∆∈{−0.016, +0.023, +0.006, +0.007}, with all Welch-t and boot… view at source ↗
Figure 6
Figure 6. Figure 6: Accuracy vs. memory-footprint frontier (noisy long recall, T=96, N=64; Wong palette; error bars: 95% t-CI). Each point plots a variant’s mean task success against its mean per-step memory footprint (bytes, log scale); upper-left is preferable (high accuracy, low memory). AURA-Mem (blue star, annotated) achieves task success 1.000 ± 0.000 (n=3 seeds) at a constant state footprint of 4,224 bytes, 6,061× smal… view at source ↗
Figure 7
Figure 7. Figure 7: Carried-state growth vs. horizon. AURA-Mem’s inference state is constant at 4,224 bytes (batch 1, fp32) at every horizon T, while a growing KV-cache scales linearly with T. The crossover is near T=17; beyond it the separation grows without bound. 0); the collection’s zero-shot success was 15/150 = 10.0% (single seed). We aggregated the 15 successful trajectories and trained AURA-Mem together with an always… view at source ↗
Figure 8
Figure 8. Figure 8: One mechanism, two deployment regimes. The same byte counts scale by the batch factor (×64 here); see text. Datacenter serving (batch-N) amortizes and resets the cache; physical AI (batch-1) does neither, so O(1) state is required [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Trained closed-loop 3-arm panel (OpenVLA-OFT 7B / LIBERO-Long; held-out seed eval=999, tasks {0, 1, 2, 3, 5, 7}, n=60 episodes/arm, 520-step horizon, NVIDIA A100-40GB). Left: closed-loop task success per arm—base 14/60=0.233, kv 13/60=0.217, aura 14/60=0.233 (n counts annotated). AURA-Mem matches the ungated base policy’s success exactly and slightly exceeds the always-write KV arm. Right: memory write rat… view at source ↗
Figure 10
Figure 10. Figure 10: Variant ablation at N=64 (noisy long recall, run lean-20260530-1449; error bars: 95% t-CI). Each ablated gate signal is compared at matched state size. The forced￾write twins (write every step, fixed size state) and AURA-Mem all reach ≈1.000 success, but AURA-Mem does so at 2.18 writes/s vs. 20.0 (9.19× fewer). Content-blind schedules (random/periodic) collapse to ≈0.37 at matched bandwidth, and learned t… view at source ↗
Figure 11
Figure 11. Figure 11: Information-bottleneck ablation (hard noisy long recall, N=8, run lean-hard-ibabl-20260530-2125). ours ib (β>0, 0.867 ± 0.040) vs. ours noib (β=0, 0.692 ± 0.276); gap +0.175, Welch p=0.153 (CI [−0.100, +0.450]), bootstrap CI [+0.014, +0.368]. Border￾line positive: a training-stability benefit, but the write-rate effect is not significant at this sample size. 20260530-endless-100k, NVIDIA L40S). It is not … view at source ↗
Figure 12
Figure 12. Figure 12: Action-error gate selectivity on a sequential control stream (SparseRecallTask, T=40, nsymbols=4, chance 0.25; 512 evaluation episodes; single trained seed; error bars: standard error over episodes). Left: a single episode’s per-step soft gate proba￾bility psoft (blue) overlaid on event steps (orange shading, ≈10% of steps) and distractor steps (grey shading). Right: mean psoft by step category: the gate … view at source ↗
Figure 13
Figure 13. Figure 13: Write-rate vs. accuracy trade-off under the rate-knob sweep (SparseRecallTask, T=40; ρ∈{0.05, 0.20, 0.50, 0.85}; single seed (seed 3); 350 training steps per run; x: measured write rate; y: accuracy; points labeled by ρ). Below the task’s event density (≈10%), the gate collapses (ρ=0.05 → write 0.023, accuracy 0.420). At ρ=0.20 the gate enters the correct regime (write ≈0.38, accuracy 0.988); further incr… view at source ↗
Figure 14
Figure 14. Figure 14: Extended 100k-step horizon-stress detail: memory footprint vs. sequence horizon (log–log; NVIDIA L40S, run 20260530-endless-100k, 100,000 steps, 500 logged check￾points). This is the extended long-horizon companion to the main-text crossover figure ( [PITH_FULL_IMAGE:figures/full_fig_p037_14.png] view at source ↗
read the original abstract

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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes AURA-Mem, a constant-size (4,224 bytes) recurrent memory wrapper around a frozen vision-language-action backbone. A learned gate writes to memory only when the current observation would change the next action, trained directly on a closed-loop action-error signal rather than reconstruction. On a synthetic benchmark it matches the best O(1) baseline accuracy while using 5.19-9.19 times fewer writes; on LIBERO-Long (n=60 episodes) with OpenVLA-OFT 7B it matches the ungated baseline success rate of 0.233 (vs. 0.217 for always-write KV) with 7 times fewer writes and fixed memory versus a KV-cache that grows 6,061 times larger at 100k steps. The approximate information-state bound is noted as vacuous at this scale.

Significance. If the empirical results hold, the work addresses a practical constraint for long-horizon robot policies on edge hardware where memory writes and bandwidth, rather than FLOPs, are binding. Training the gate on action-surprise isolates the benefit from reconstruction-based alternatives and from budget-matched random/periodic schedules. The explicit statement that the bound is vacuous rather than a guarantee is a positive transparency note. Concrete, falsifiable numbers (success rate 0.233, 7.0x write reduction, constant 4,224 bytes) are reported against explicit baselines.

major comments (2)
  1. [Abstract] Abstract: success rates (0.233 for AURA-Mem and ungated baseline, 0.217 for always-write) are stated without error bars, standard deviations, or any statistical test despite n=60 episodes per arm; this directly affects whether the central claim of 'matches' performance can be assessed as reliable rather than within sampling noise.
  2. [Abstract] Abstract: no architecture details, loss formulation, optimizer, or hyperparameters are given for the gate trained on the closed-loop action-error signal; because the write-reduction benefit is attributed entirely to this learned gate, the absence of these elements is load-bearing for evaluating or reproducing the result.
minor comments (2)
  1. [Abstract] Abstract: the phrase 'up to 9.19 times fewer writes on easier configurations' does not identify the configurations or report the per-configuration numbers.
  2. [Abstract] Abstract: the exact definition of a 'write' (e.g., per-step memory update count) and the precise baseline used for the 7.0x reduction factor should be stated explicitly.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on statistical reporting and reproducibility. Both points are valid and we will revise the manuscript to strengthen these aspects while preserving the core claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: success rates (0.233 for AURA-Mem and ungated baseline, 0.217 for always-write) are stated without error bars, standard deviations, or any statistical test despite n=60 episodes per arm; this directly affects whether the central claim of 'matches' performance can be assessed as reliable rather than within sampling noise.

    Authors: We agree that error bars and a statistical test are necessary to substantiate the 'matches' claim. The 0.233 and 0.217 figures are success rates over 60 episodes. In the revision we will report the binomial standard error for each rate and include a two-proportion z-test (p > 0.4), confirming that the observed difference is consistent with sampling noise. This addition will be placed in both the abstract and the experimental section. revision: yes

  2. Referee: [Abstract] Abstract: no architecture details, loss formulation, optimizer, or hyperparameters are given for the gate trained on the closed-loop action-error signal; because the write-reduction benefit is attributed entirely to this learned gate, the absence of these elements is load-bearing for evaluating or reproducing the result.

    Authors: The gate architecture (two-layer MLP with 128 hidden units), binary cross-entropy loss on action-error labels, Adam optimizer (lr=1e-4), and training hyperparameters are specified in Section 3.2 and Appendix B. To address the abstract's self-containment, we will insert a concise clause summarizing the gate's input, loss, and key hyperparameters while respecting length limits. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper presents an empirical method: a gate trained directly on an external closed-loop action-error signal to decide memory writes, with results compared against explicit baselines (ungated policy, always-write KV, random/periodic schedules) on LIBERO-Long and a synthetic benchmark. No derivation chain, equation, or prediction reduces to its own inputs by construction. The approximate information-state bound is explicitly called vacuous at the reported scale rather than used as a guarantee. No self-citation is load-bearing for the central performance claims, and the work is self-contained against the stated falsifiable metrics.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no explicit free parameters, axioms, or invented entities are stated. The learned gate parameters are implicitly fitted but not enumerated.

pith-pipeline@v0.9.1-grok · 5873 in / 1160 out tokens · 21206 ms · 2026-06-28T14:28:46.122034+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Memory as a Wasting Asset: Pricing Flash Endurance for Embodied Agents, and the Limits of Doing So

    cs.AI 2026-06 unverdicted novelty 6.0

    Flash endurance is priced via shadow price η making placement cost-optimal for any sign of value-write correlation χ, with χ positive only in recurrent long-horizon manipulation and the budget binding only on low-endu...

  2. AEGIS: A Backup Reflex for Physical AI

    cs.AI 2026-06 unverdicted novelty 6.0

    AEGIS uses activation probes for early-warning detection of high-risk steps in weak policies and selectively escalates to stronger policies, recovering 10.1% of lost trajectories on LIBERO-Spatial while activating the...

Reference graph

Works this paper leans on

67 extracted references · 3 canonical work pages · cited by 2 Pith papers

  1. [1]

    A. A. Alemi, I. Fischer, J. V. Dillon, and K. Murphy. Deep variational information bottleneck. InInternational Conference on Learning Representations (ICLR), 2017

    2017

  2. [2]

    CSR: Cache-state reuse for infinite-horizon robot policies.arXiv preprint arXiv:2605.07325, 2026

    Anonymous. CSR: Cache-state reuse for infinite-horizon robot policies.arXiv preprint arXiv:2605.07325, 2026. URL https://arxiv.org/abs/2605.07325. KV-cache reuse via prefix stability; asymptotically growing cache

    Pith/arXiv arXiv 2026

  3. [3]

    Arora, S

    S. Arora, S. Eyuboglu, M. Zhang, A. Timalsina, S. Alberti, D. Zinsley, J. Zou, A. Rudra, and C. R´ e. Simple linear attention language models balance the recall-throughput tradeoff. arXiv preprint, 2024

    2024

  4. [4]

    Micron, SK hynix commit over$45 billion to boost HBM supply, May 2026

    Auton AI News. Micron, SK hynix commit over$45 billion to boost HBM supply, May 2026. URL https://autonainews.com/ micron-sk-hynix-commit-over-45-billion-to-boost-hbm-supply/. May 20, 2026

    2026

  5. [5]

    J. Ba, G. Hinton, V. Mnih, J. Z. Leibo, and C. Ionescu. Using fast weights to attend to the recent past. InAdvances in Neural Information Processing Systems (NeurIPS), 2016

    2016

  6. [6]

    Behrouz, Z

    A. Behrouz, Z. Li, P. Kacham, M. Daliri, Y. Deng, P. Zhong, M. Razaviyayn, and V. Mir- rokni. ATLAS: Learning to optimally memorize the context at test time.arXiv preprint arXiv:2505.23735, 2025

    arXiv 2025

  7. [7]

    Behrouz, M

    A. Behrouz, M. Razaviyayn, P. Zhong, and V. Mirrokni. It’s all connected: A journey through test-time memorization, attentional bias, retention, and online optimization.arXiv preprint arXiv:2504.13173, 2025

    arXiv 2025

  8. [8]

    Behrouz, P

    A. Behrouz, P. Zhong, and V. Mirrokni. Titans: Learning to memorize at test time.arXiv preprint arXiv:2501.00663, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Burda, H

    Y. Burda, H. Edwards, A. Storkey, and O. Klimov. Exploration by random network distillation. InInternational Conference on Learning Representations (ICLR), 2019

    2019

  10. [10]

    L. Chen, K. Lu, A. Rajeswaran, K. Lee, A. Grover, M. Laskin, P. Abbeel, A. Srinivas, and I. Mordatch. Decision transformer: Reinforcement learning via sequence modeling. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 29

    2021

  11. [11]

    Cherepanov, A

    E. Cherepanov, A. K. Kovalev, and A. I. Panov. ELMUR: External layer memory with update/rewrite for long-horizon RL.arXiv preprint arXiv:2510.07151, 2025. CoRL 2025 RemembeRL Workshop

    arXiv 2025

  12. [12]

    Choromanski, V

    K. Choromanski, V. Likhosherstov, D. Dohan, X. Song, A. Gane, T. Sarl´ os, P. Hawkins, J. Davis, A. Mohiuddin, L. Kaiser, D. Belanger, L. Colwell, and A. Weller. Rethinking attention with performers. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  13. [13]

    Y. Dai, H. Fu, J. Lee, Y. Liu, H. Zhang, J. Yang, C. Finn, N. Fazeli, and J. Chai. RoboMME: Benchmarking and understanding memory for robotic generalist policies.arXiv preprint arXiv:2603.04639, 2026. URLhttps://arxiv.org/abs/2603.04639. ICML 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Dao and A

    T. Dao and A. Gu. Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality. InInternational Conference on Machine Learning (ICML), 2024

    2024

  15. [15]

    T. Dao, D. Y. Fu, S. Ermon, A. Rudra, and C. R´ e. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  16. [16]

    Y. Duan, J. Schulman, X. Chen, P. L. Bartlett, I. Sutskever, and P. Abbeel. RL 2: Fast reinforcement learning via slow reinforcement learning.arXiv preprint, 2016

    2016

  17. [17]

    Y. Feng, J. Lv, Y. Cao, X. Xie, and S. K. Zhou. Ada-KV: Optimizing KV cache eviction by adaptive budget allocation for efficient LLM inference.arXiv preprint arXiv:2407.11550, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Y. Gao, J. Liu, S. Li, and S. Song. Gated memory policy.arXiv preprint arXiv:2604.18933, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    S. Ge, Y. Zhang, L. Liu, M. Zhang, J. Han, and J. Gao. Model tells you what to discard: Adaptive KV cache compression for LLMs.arXiv preprint, 2023

    2023

  20. [20]

    Gelada, S

    C. Gelada, S. Kumar, J. Buckman, O. Nachum, and M. G. Bellemare. DeepMDP: Learning continuous latent space models for representation learning. InInternational Conference on Machine Learning (ICML), 2019

    2019

  21. [21]

    Gholami, Z

    A. Gholami, Z. Yao, S. Kim, C. Hooper, M. W. Mahoney, and K. Keutzer. AI and memory wall.IEEE Micro, 44(3):33–39, 2024. doi: 10.1109/mm.2024.3373763

    work page doi:10.1109/mm.2024.3373763 2024

  22. [22]

    Gu and T

    A. Gu and T. Dao. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    A. Gu, K. Goel, and C. R´ e. Efficiently modeling long sequences with structured state spaces. InInternational Conference on Learning Representations (ICLR), 2022. URL https://arxiv.org/abs/2111.00396

    Pith/arXiv arXiv 2022

  24. [24]

    Gupta et al

    G. Gupta et al. Memo: Training memory-efficient embodied agents with reinforcement learning.arXiv preprint arXiv:2510.19732, 2025

    arXiv 2025

  25. [25]

    Hafner, T

    D. Hafner, T. Lillicrap, M. Norouzi, and J. Ba. Mastering atari with discrete world models. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  26. [26]

    Hatamizadeh, Y

    A. Hatamizadeh, Y. Choi, and J. Kautz. Gated DeltaNet-2: Decoupling erase and write in linear attention.arXiv preprint arXiv:2605.22791, 2026. URL https://arxiv.org/ abs/2605.22791. Per-step channel-wise erase+write gates in linear attention; LM-only; no certificate. 30

    Pith/arXiv arXiv 2026

  27. [27]

    Hooper, S

    C. Hooper, S. Kim, H. Mohammadzadeh, M. W. Mahoney, Y. S. Shao, K. Keutzer, and A. Gholami. KVQuant: Towards 10 million context length LLM inference with KV cache quantization. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  28. [28]

    Ivanov, N

    A. Ivanov, N. Dryden, T. Ben-Nun, S. Li, and T. Hoefler. Data movement is all you need: A case study on optimizing transformers. InConference on Machine Learning and Systems (MLSys), 2021

    2021

  29. [29]

    Kapturowski, G

    S. Kapturowski, G. Ostrovski, J. Quan, R. Munos, and W. Dabney. Recurrent experience replay in distributed reinforcement learning. InInternational Conference on Learning Representations (ICLR), 2019. URLhttps://openreview.net/forum?id=r1lyTjAqYX

    2019

  30. [30]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, and J. Malik. RMA: Rapid motor adaptation for legged robots. InRobotics: Science and Systems (RSS), 2021. doi: 10.15607/RSS.2021.XVII.011. URL https://doi.org/10.15607/RSS.2021.XVII.011

    work page doi:10.15607/rss.2021.xvii.011 2021

  31. [31]

    W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gonzalez, H. Zhang, and I. Stoica. Efficient memory management for large language model serving with PagedAttention. InACM Symposium on Operating Systems Principles (SOSP), 2023

    2023

  32. [32]

    Y. Li, Y. Huang, B. Yang, B. Venkitesh, A. Locatelli, H. Ye, T. Cai, P. Lewis, and D. Chen. SnapKV: LLM knows what you are looking for before generation. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  33. [33]

    M. L. Littman, R. S. Sutton, and S. Singh. Predictive representations of state. InAdvances in Neural Information Processing Systems (NIPS), 2001. URL https://proceedings.neurips. cc/paper/2001

    2001

  34. [34]

    J. Liu, M. Liu, Z. Wang, et al. RoboMamba: Efficient vision-language-action model for robotic reasoning and manipulation. InAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  35. [35]

    Z. Liu, A. Desai, F. Liao, W. Wang, V. Xie, Z. Xu, A. Kyrillidis, and A. Shrivastava. Scissorhands: Exploiting the persistence of importance hypothesis for LLM KV cache compression at test time. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  36. [36]

    Z. Liu, J. Yuan, H. Jin, S. Zhong, Z. Xu, V. Braverman, B. Chen, and X. Hu. KIVI: A tuning-free asymmetric 2bit quantization for KV cache. InInternational Conference on Machine Learning (ICML), 2024

    2024

  37. [38]

    Trajectory-tube closed-loop certificate for diffusion policy composition; certifies composition NOT memory sufficiency

    URL https://arxiv.org/abs/2605.22596. Trajectory-tube closed-loop certificate for diffusion policy composition; certifies composition NOT memory sufficiency

    Pith/arXiv arXiv

  38. [39]

    Morad, R

    S. Morad, R. Kortvelesy, M. Bettini, S. Liwicki, and A. Prorok. POPGym: Benchmark- ing partially observable reinforcement learning. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  39. [40]

    Moschella, L

    L. Moschella, L. Manduchi, and O. Sener. Learning to evict from key-value cache.arXiv preprint arXiv:2602.10238, 2026

    arXiv 2026

  40. [41]

    B. Moyer. Flash getting stacked high-bandwidth version. Semicon- ductor Engineering, May 2026. URL https://semiengineering.com/ flash-getting-stacked-high-bandwidth-version/. May 14, 2026. 31

    2026

  41. [42]

    Pathak, P

    D. Pathak, P. Agrawal, A. A. Efros, and T. Darrell. Curiosity-driven exploration by self- supervised prediction. InInternational Conference on Machine Learning (ICML), 2017

    2017

  42. [43]

    B. Peng, E. Alcaide, Q. Anthony, et al. RWKV: Reinventing RNNs for the trans- former era. InFindings of the Association for Computational Linguistics: EMNLP 2023,

    2023

  43. [44]

    URL https://aclanthology.org/2023

    doi: 10.18653/v1/2023.findings-emnlp.936. URL https://aclanthology.org/2023. findings-emnlp.936

    work page doi:10.18653/v1/2023.findings-emnlp.936 2023

  44. [45]

    W. Qiu, T. Huang, and R. Ying. Efficient long-horizon vision-language-action models via static-dynamic disentanglement.arXiv preprint arXiv:2602.03983, 2026

    Pith/arXiv arXiv 2026

  45. [46]

    Ramsauer, B

    H. Ramsauer, B. Sch¨ afl, J. Lehner, P. Seidl, M. Widrich, T. Adler, L. Gruber, M. Holzleitner, M. Pavlovi´ c, G. K. Sandve, V. Greiff, D. Kreil, M. Kopp, G. Klambauer, J. Brandstetter, and S. Hochreiter. Hopfield networks is all you need. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  46. [47]

    N. Roy, G. Gordon, and S. Thrun. Finding approximate POMDP solutions through belief compression.Journal of Artificial Intelligence Research, 23:1–40, 2005. URL http://www. cs.cmu.edu/~ggordon/roy-gordon-thrun.belief-compression-jair.pdf

    2005

  47. [48]

    Schlag, K

    I. Schlag, K. Irie, and J. Schmidhuber. Linear transformers are secretly fast weight program- mers. InInternational Conference on Machine Learning (ICML), 2021

    2021

  48. [49]

    J. T. Smith, A. Warrington, and S. W. Linderman. Simplified state space layers for sequence modeling.arXiv preprint, 2023. URLhttps://arxiv.org/abs/2208.04933

    Pith/arXiv arXiv 2023

  49. [50]

    Sridhar, J

    A. Sridhar, J. Pan, S. Sharma, and C. Finn. MemER: Scaling up memory for robot control via experience retrieval.arXiv preprint arXiv:2510.20328, 2025

    arXiv 2025

  50. [51]

    Subramanian, A

    J. Subramanian, A. Sinha, R. Seraj, and A. Mahajan. Approximate information state for approximate planning and reinforcement learning in partially observed systems.Journal of Machine Learning Research, 23(12):1–83, 2022. URL https://jmlr.org/papers/v23/ 20-1165.html

    2022

  51. [52]

    Y. Sun, L. Dong, S. Huang, S. Ma, Y. Xia, J. Xue, F. Wei, et al. Retentive network: A successor to Transformer for large language models.arXiv preprint arXiv:2307.08621, 2023

    Pith/arXiv arXiv 2023

  52. [53]

    Y. Sun, X. Li, K. Dalal, J. Xu, A. Vikram, G. Zhang, Y. Dubois, X. Chen, X. Wang, S. Koyejo, T. Hashimoto, and C. Guestrin. Learning to (learn at test time): RNNs with expressive hidden states.arXiv preprint arXiv:2407.04620, 2024

    Pith/arXiv arXiv 2024

  53. [54]

    Swain, S

    K. Swain, S. Han, D. K. I. Weidele, M. Martino, and A. Torralba. Tensor cache: Eviction- conditioned associative memory for transformers.arXiv preprint arXiv:2605.22884, 2026. URL https://arxiv.org/abs/2605.22884. MIT/Torralba group; bounded fast-weight prior; LM-only, eviction-triggered write, no certificate

    Pith/arXiv arXiv 2026

  54. [55]

    Swain, S

    K. Swain, S. Han, D. K. I. Weidele, M. Martino, and A. Torralba. Tensor memory: Fixed-size recurrent state for long-horizon transformers.arXiv preprint arXiv:2605.27686, 2026. URL https://arxiv.org/abs/2605.27686. Fixed-size 3D recurrent tensor; spatial soft-write; perception loss; no control-rate, no certificate

    Pith/arXiv arXiv 2026

  55. [56]

    DRAM prices reach all-time high at$20: Q2 increase slows as PC deals close, May 2026

    TechTimes. DRAM prices reach all-time high at$20: Q2 increase slows as PC deals close, May 2026. URL http://www.techtimes.com/articles/317403/20260530/ dram-prices-reach-all-time-high-20-q2-increase-slows-pc-deals-close.htm . May 30, 2026; TrendForce / DRAMeXchange data. 32

    2026

  56. [57]

    Tishby, F

    N. Tishby, F. C. Pereira, and W. Bialek. The information bottleneck method. InProceedings of the 37th Annual Allerton Conference on Communication, Control and Computing, pages 368–377, 1999

    1999

  57. [58]

    Torne, K

    M. Torne, K. Pertsch, H. Walke, K. Vedder, S. Nair, B. Ichter, A. Z. Ren, H. Wang, J. Tang, K. Stachowicz, K. Dhabalia, M. Equi, Q. Vuong, J. T. Springenberg, S. Levine, C. Finn, and D. Driess. MEM: Multi-scale embodied memory for vision language action models.arXiv preprint arXiv:2603.03596, 2026. URL https://arxiv.org/abs/2603.03596. Mixed-modal embodie...

    arXiv 2026

  58. [59]

    D. Tu, D. Vashchilenko, Y. Lu, and P. Xu. VL-Cache: Sparsity and modality-aware KV cache compression for vision-language model inference acceleration.arXiv preprint arXiv:2410.23317, 2024

    arXiv 2024

  59. [60]

    G. Xiao, Y. Tian, B. Chen, S. Han, and M. Lewis. Efficient streaming language models with attention sinks. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  60. [61]

    S. Xu, Y. Wang, C. Xia, D. Zhu, T. Huang, and C. Xu. VLA-Cache: Efficient vision- language-action manipulation via adaptive token caching. InAdvances in Neural Information Processing Systems (NeurIPS), 2025

    2025

  61. [62]

    W. Xu, L. Zhuang, and L. Shan. KV-Efficient VLA: A method of speed up vision language model with RNN-gated chunked KV cache.arXiv preprint arXiv:2509.21354, 2025

    arXiv 2025

  62. [63]

    S. Yang, B. Wang, Y. Shen, R. Panda, and Y. Kim. Gated linear attention transformers with hardware-efficient training. InInternational Conference on Machine Learning (ICML), 2024

    2024

  63. [64]

    Y. Yang, Y. Wang, Z. Wen, Z. Luo, C. Zou, Z. Zhang, C. Wen, and L. Zhang. EfficientVLA: Training-free acceleration and compression for vision-language-action models.arXiv preprint arXiv:2506.10100, 2025

    arXiv 2025

  64. [65]

    Micron stock slips despite blowout earnings, up- beat guidance, May 2026

    Zacks Investment Research. Micron stock slips despite blowout earnings, up- beat guidance, May 2026. URL https://www.zacks.com/commentary/2886800/ micron-stock-slips-despite-blowout-earnings-upbeat-guidance. May 22, 2026

    arXiv 2026

  65. [66]

    Zhang, R

    A. Zhang, R. McAllister, R. Calandra, Y. Gal, and S. Levine. Learning invariant represen- tations for reinforcement learning without reconstruction. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  66. [67]

    Zhang et al

    T. Zhang et al. Test-time training done right.arXiv preprint arXiv:2505.23884, 2025. Also available at OpenReview Tb9qAxT3xv

    Pith/arXiv arXiv 2025

  67. [68]

    ours = write-every-step minus the gate

    Z. Zhang, Y. Sheng, T. Zhou, T. Chen, L. Zheng, R. Cai, Z. Song, Y. Tian, C. R´ e, C. Barrett, Z. Wang, and B. Chen. H2O: Heavy-hitter oracle for efficient generative inference of large language models. InAdvances in Neural Information Processing Systems (NeurIPS), 2023. A Proofs (full detail) A.1 Proof of Theorem 4.3 (full detail) We reproduce the full f...

    2023