pith. sign in

arxiv: 2606.18144 · v1 · pith:5YIIL3EUnew · submitted 2026-06-16 · 💻 cs.AI · cs.CY· cs.LG· cs.RO

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

Josef Liyanjun Chen This is my paper

Pith reviewed 2026-06-27 00:53 UTC · model grok-4.3

classification 💻 cs.AI cs.CYcs.LGcs.RO
keywords flash endurancerobot memory managementshadow pricewear-aware placementembodied agentsvalue-write associationmemory hierarchyendurance budget
0
0 comments X

The pith

A single endurance shadow price turns robot memory placement into a cost-optimal threshold rule across RAM, flash, and cloud.

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

The paper treats a robot's flash memory endurance as a limited, non-renewable resource that gets used up with each write. It introduces an endurance shadow price to decide which memories to keep on which storage tier. This pricing makes the placement decision cost-minimal no matter whether high-value memories are written often or rarely. Measurements on real robot logs show the value-write link is positive only in certain long-horizon tasks. On cheap flash chips the budget binds, but a simple price rule matches learned controllers because value does not change with storage tier.

Core claim

The central claim is that pricing flash endurance with a single shadow price η makes the cost-minimizing placement of memories across a RAM/on-board NVM/cloud hierarchy a threshold rule on a wear-augmented per-byte index. This index remains cost-optimal for any sign of the value-write association χ. Only when χ is positive does the optimum become non-monotone, moving the robot's most valuable memories off its flash. The endurance budget is dormant on premium TLC flash but binding on commodity QLC/eMMC. Where it binds, price-based routing ties learned wear-aware controllers on task value because realized value is tier-invariant.

What carries the argument

the endurance shadow price η that converts memory placement into a threshold on a wear-augmented per-byte index

If this is right

  • The placement rule works regardless of whether valuable data is written more frequently.
  • When the association is positive, the most valuable memories are evicted from flash.
  • The budget only constrains decisions on lower-endurance commodity flash.
  • Price-based routing performs as well as learned wear-aware methods on task value.
  • The non-monotone optimum remains untested in data.

Where Pith is reading between the lines

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

  • If the tier-invariance of value holds, then task performance depends only on access speed, not on where data lives.
  • This suggests that for cheaper robots, endurance pricing could extend device life without hurting capabilities.
  • Testing the non-monotone case would require logs from long-horizon recurrent tasks with varying memory values.

Load-bearing premise

Value realized by a memory stays the same whether it sits in RAM, on flash, or in the cloud, so only the rental cost and endurance matter for decisions.

What would settle it

Observing whether the non-monotone placement optimum actually appears in robot logs from recurrent manipulation tasks, or measuring if a wear-aware controller ever beats the price rule on realized task value.

Figures

Figures reproduced from arXiv: 2606.18144 by Josef Liyanjun Chen.

Figure 1
Figure 1. Figure 1: The core idea in one picture. On-board NAND endurance is a non-renewable stock: each persisted write spends one of a few thousand program/erase cycles and is gone, whereas RAM capacity refills every period. Every retained memory therefore faces a priced choice—keep in RAM, persist to flash (spending an erase cycle), offload to cloud, or forget—governed by a single endurance rent η. No fielded embodied-memo… view at source ↗
Figure 2
Figure 2. Figure 2: Research-program-arc banner. AURA decides when to write; this paper decides where memory lives; the economic layer prices what it is worth (the endurance rent η). stock is scarce, and scarcity is a regime, not a given—dormant on premium TLC at datasheet prices, binding on the commodity QLC/eMMC cheaper edge robots run (section 5.3). The non-monotone refinement (Proposition 2) needs one further primitive, a… view at source ↗
Figure 3
Figure 3. Figure 3: Every headline claim, by epistemic tier. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Wear phase diagram (illustrative; model-derived). [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Calibrated economic primitives. Left: the wear+rent break-even ($/GB-day) at Low/Base/High NAND (≈ $4.8 × 10−5 at base). Right: the Jorgenson user cost of one erase, decom￾posed into datasheet cash wear plus the +5.0% endurance markup (mˆ ). The persisted item is worth M = 24.3× its per-step value (half-life 21.7 steps). Phase 0 value–write gate χˆ, ρs on real robot logs (≤ $25) Phase 1 value labeling Smol… view at source ↗
Figure 6
Figure 6. Figure 6: Experiment-architecture pipeline. Five pre-specified stages, each with a published kill criterion (red): the $25 Phase-0 gate decides whether the non-monotone branch is admissible before any training spend; Phase 1 labels item value via action counterfactuals; Phase 2 trains the 3.15M-parameter placement controller; Phase 3 runs the cost-matched McNemar/Holm battery; the final stage closes the loop in the … view at source ↗
Figure 7
Figure 7. Figure 7: The value→write coupling’s sign is regime-dependent. Canonical χˆ with 95% scene￾clustered CIs: positive and CI-excluding-zero on recurrent LIBERO-Long (SmolVLA-0.5B), significantly negative on non-recurrent DROID (post-hoc), and CI-straddling on the OpenVLA-7B scale-stress arm, whose cross-backbone agreement (ρs = 0.05 ≪ 0.6) makes its sign uninterpretable. regime-difference signal, not a confirmation. Th… view at source ↗
Figure 8
Figure 8. Figure 8: Recurrence-driven vs. churn-driven write-intensity: the mechanism behind the sign. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Recurrence dose-response replicates at full power. [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The endurance budget binds on commodity storage. [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: When wear-aware placement beats endurance-blind routing. [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The recurrence–dispersion tension. A wear-aware win needs both χ > 0 and high write￾dispersion CV(w), but they are anti-correlated across all three write-generating mechanisms: recurrence (manipulation) gives χ > 0 at low dispersion; churn (teleoperation) and navigation give high dispersion but χ < 0. Four in-house workloads (round/plus/square/diamond) and nine external robot workloads (stars) from an ≈ 2… view at source ↗
Figure 13
Figure 13. Figure 13: Controller robustness and the metric-gaming finding. [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Price comparative-statics fan over the unified Low/Base/High band (real re-solved outputs): the equilibrium rent ηsim declines monotonically across the band (left) while the population NVM share σN stays budget-pinned at ≈ 0.29–0.30 (right), with the break-even v ⋆ BE = 0.91 annotated. 6 Discussion What the priced model buys. The central object is the endurance shadow price η: it fixes the persist/evict b… view at source ↗
Figure 15
Figure 15. Figure 15: When wear-aware placement switches on. The lever needs both a binding endurance budget (η > 0, the right of the map) and a positive value–write coupling (χ > 0, the top): only the upper-right cell activates the non-monotone optimum. Our measured datasets sit in the endurance-abundant band as run (LIBERO-Long recurrent χ > 0; DROID churn-driven χ < 0), but the binding (right) column is live today on the co… view at source ↗
read the original abstract

A robot's flash endurance is a non-renewable stock: every persisted write spends one of a few thousand program/erase cycles and never refills, yet no fielded robot memory system prices which memories are worth an erase cycle. We treat embodied memory as depreciating capital and price that stock with a single endurance shadow price $\eta$, which makes cost-minimizing placement across a RAM / on-board NVM / cloud hierarchy a threshold in a wear-augmented per-byte index. The index is cost-optimal whatever the sign of the value-write association $\chi$; only when $\chi > 0$ does the optimum turn non-monotone, sending a robot's most valuable memories off its flash. The pivot is thus empirical, and we measure $\chi$ on real robot logs at a pre-specified gate: its sign is a property of the deployment regime -- positive on recurrent long-horizon manipulation ($\hat{\chi} \approx +1.0 \times 10^{-3}$, replicated at full power), null on a shorter-horizon suite, and negative on non-recurrent teleoperation. Two boundaries scope the result. The endurance budget is dormant on premium 3,000-P/E TLC at datasheet prices and binding on the commodity QLC/eMMC ($\sim$1,000 P/E) that cheaper edge robots run. And where it binds, a learned wear-aware controller only ties price-based routing on task value, because realized value is tier-invariant across RAM, NVM, and cloud: the rent governs device lifetime and cost, not task performance. Whether wear-aware placement improves task value remains open -- $\chi$ is measured against a value proxy, and the non-monotone optimum, while proven, is not yet observed in data.

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 / 1 minor

Summary. The paper claims that flash endurance can be priced via a shadow price η to derive a wear-augmented per-byte index for cost-minimizing placement of memories across RAM/NVM/cloud in embodied agents. The index is asserted to be optimal for any sign of the value-write association χ; non-monotonicity (and thus eviction of high-value memories) arises only when χ > 0. Empirically, χ is measured from external robot logs at a pre-specified gate, yielding positive values for recurrent long-horizon manipulation, null or negative otherwise. Endurance is dormant at premium 3000 P/E TLC datasheet prices but binding at commodity ~1000 P/E QLC/eMMC levels; a learned wear-aware controller only ties price-based routing because realized value is tier-invariant, so rent (not task performance) governs lifetime.

Significance. If the central derivation and empirical sign-of-χ result hold, the work supplies a first-principles cost-minimization framework for non-renewable memory stock in robots, with clear hardware-scope boundaries and an explicit statement that the non-monotone optimum remains unobserved. Credit is due for deriving the placement rule from external logs rather than internal fitting (avoiding tautology) and for the transparent acknowledgment that wear-aware gains on task value are untested.

major comments (2)
  1. [Empirical measurement of χ (abstract and results section)] The empirical pivot on χ (measured at a pre-specified gate on robot logs) is load-bearing for the claim that its sign is a property of the deployment regime and for the conclusion that endurance binds only on low-P/E media. The manuscript must supply the exact gate definition, error bars on the reported χ̂ ≈ +1.0 imes 10^{-3}, exclusion rules, and replication protocol (including the 'full power' replication) to allow assessment of robustness; these details are not visible even in the abstract.
  2. [Derivation of non-monotonicity and discussion of unobserved optimum] The non-monotone optimum is derived for χ > 0 but explicitly stated to be unobserved in data. Because this is the distinctive practical implication (most valuable memories sent off flash), the manuscript should either provide a concrete test or bound the conditions under which the non-monotonicity would appear, rather than leaving it as an open empirical gap.
minor comments (1)
  1. Notation for the endurance shadow price η and the association χ is introduced without an early consolidated table of symbols; a short notation table would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. We address each major comment below. We will expand the empirical details on χ measurement in the revised manuscript. For the non-monotonic optimum, we will add explicit bounds on the conditions for its appearance while acknowledging that a direct empirical test lies outside the current datasets.

read point-by-point responses

  1. Referee: [Empirical measurement of χ (abstract and results section)] The empirical pivot on χ (measured at a pre-specified gate on robot logs) is load-bearing for the claim that its sign is a property of the deployment regime and for the conclusion that endurance binds only on low-P/E media. The manuscript must supply the exact gate definition, error bars on the reported χ̂ ≈ +1.0 times 10^{-3}, exclusion rules, and replication protocol (including the 'full power' replication) to allow assessment of robustness; these details are not visible even in the abstract.

    Authors: We agree that these methodological details are necessary to evaluate robustness. In the revised manuscript we will add, in the results section (with a cross-reference from the abstract), the precise definition of the pre-specified gate used on the robot logs, the formula and numerical error bars for χ̂, the exclusion criteria applied to log entries, and the complete replication protocol including the full-power replication. These additions will be presented without altering the reported sign or magnitude of χ. revision: yes

  2. Referee: [Derivation of non-monotonicity and discussion of unobserved optimum] The non-monotone optimum is derived for χ > 0 but explicitly stated to be unobserved in data. Because this is the distinctive practical implication (most valuable memories sent off flash), the manuscript should either provide a concrete test or bound the conditions under which the non-monotonicity would appear, rather than leaving it as an open empirical gap.

    Authors: The model derivation establishes that non-monotonicity arises exclusively when χ > 0. We accept that the current datasets do not exhibit the non-monotone placement regime itself. In revision we will insert an explicit bounding analysis (new subsection) that states the minimum χ magnitude and endurance-budget tightness required for the non-monotone optimum to appear, derived directly from the closed-form threshold condition. A controlled empirical demonstration of the non-monotone regime would require new robot experiments with higher χ or lower-P/E hardware; we will note this as future work rather than claim it is resolved here. revision: partial

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The derivation obtains the wear-augmented per-byte index directly from cost minimization across the RAM/NVM/cloud hierarchy once the endurance shadow price η is introduced; the optimality statement for any sign of χ follows mathematically from that construction and does not reduce to a fitted input or self-definition. χ itself is obtained by external measurement on robot logs at a pre-specified gate, not estimated inside the model or renamed as a prediction. No self-citation chain, uniqueness theorem, or ansatz is invoked as load-bearing support. The tier-invariance assumption and the binding/non-binding endurance boundaries are stated empirical scope conditions rather than internal reductions. The central claim therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The model introduces one free parameter η (shadow price) and relies on an empirical measurement of χ; the key domain assumption is tier-invariance of realized value. No new physical entities are postulated.

free parameters (2)
  • η
    Endurance shadow price that converts each write into a cost for the placement index.
  • χ
    Value-write association measured from logs; its sign determines whether the optimum is monotone or non-monotone.
axioms (1)
  • domain assumption Realized value of a memory is invariant to the storage tier that holds it.
    Invoked to conclude that wear-aware placement affects only cost, not task performance.

pith-pipeline@v0.9.1-grok · 5865 in / 1524 out tokens · 34535 ms · 2026-06-27T00:53:56.760077+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

66 extracted references · 21 canonical work pages

  1. [1]

    Mersereau

    Daniel Adelman and Adam J. Mersereau. Relaxations of weakly coupled stochastic dynamic programs.Operations Research, 56(3):712–727, 2008. doi: 10.1287/opre.1070.0445

    work page doi:10.1287/opre.1070.0445 2008

  2. [2]

    Chapman and Hall/CRC, 1999

    Eitan Altman.Constrained Markov Decision Processes. Chapman and Hall/CRC, 1999. URL https://www-sop.inria.fr/members/Eitan.Altman/TEMP/h.pdf. 29 Table 5:Run-ID provenance for the main results. Result Run ID(s) Phase-0 value–write gatephase0-full-20260612(GATE-cov) Canonicalχmatrix / re-analysischi-reanalysis-20260612(D-012);W1+W2 battery Phase-1 labeling ...

    1999

  3. [3]

    ReMEmbR: Building and reasoning over long-horizon spatio-temporal memory for robot navigation, 2024

    Abrar Anwar, John Welsh, Joydeep Biswas, Soha Pouya, and Yan Chang. ReMEmbR: Building and reasoning over long-horizon spatio-temporal memory for robot navigation, 2024. URL https://arxiv.org/abs/2409.13682

    arXiv 2024

  4. [4]

    The financial guide to calculating the true TCO of a robotic fleet, 2025

    Balyo. The financial guide to calculating the true TCO of a robotic fleet, 2025. URL https://www.balyo.com/blog/the-financial-guide-to-calculating-the-true-tco-o f-a-robotic-fleet

    2025

  5. [5]

    A primal-dual randomized algorithm for weighted paging.Journal of the ACM, 59(4), 2012

    Nikhil Bansal, Niv Buchbinder, and Joseph (Seffi) Naor. A primal-dual randomized algorithm for weighted paging.Journal of the ACM, 59(4), 2012. doi: 10.1145/2339123.2339126. Preliminary version FOCS 2007

    work page doi:10.1145/2339123.2339126 2012

  6. [6]

    On the promise and pitfalls of optimizing embodied carbon

    Noman Bashir, David Irwin, and Prashant Shenoy. On the promise and pitfalls of optimizing embodied carbon. InProc. 2nd Workshop on Sustainable Computer Systems (HotCarbon),

  7. [7]

    URLhttps://arxiv.org/abs/2306.15816

    doi: 10.1145/3604930.3605710. URLhttps://arxiv.org/abs/2306.15816

    work page doi:10.1145/3604930.3605710

  8. [8]

    Laszlo A. Belady. A study of replacement algorithms for a virtual-storage computer.IBM Systems Journal, 5(2):78–101, 1966. doi: 10.1147/sj.52.0078

    work page doi:10.1147/sj.52.0078 1966

  9. [9]

    AEGIS: A backup reflex for physical AI, 2026

    Josef Chen. AEGIS: A backup reflex for physical AI, 2026. URLhttps://arxiv.org/abs/26 06.06660

    2026

  10. [10]

    AURA: Action-gated memory for robot policies at constant VRAM, 2026

    Josef Chen. AURA: Action-gated memory for robot policies at constant VRAM, 2026. URL https://arxiv.org/abs/2606.02775

    Pith/arXiv arXiv 2026

  11. [11]

    Memory-bound but not bandwidth-limited: The physical AI inference gap in batch-1 LLM decode, 2026

    Josef Chen. Memory-bound but not bandwidth-limited: The physical AI inference gap in batch-1 LLM decode, 2026. URL https://arxiv.org/abs/2605.30571 . Batch-1 decode across H100/A100/L40S/L4; L4 reaches 81% of analytic memory floor vs H100 27%

    Pith/arXiv arXiv 2026

  12. [12]

    Toward robust and efficient ML-based GPU caching for modern inference (lcr/laru).arXiv preprint arXiv:2509.20979, 2025

    Peng Chen, Jiaji Zhang, Hailiang Zhao, Yirong Zhang, Shenyao Chen, Jiahong Yu, Xueyan Tang, Yixuan Wang, Hao Li, Jianping Zou, Gang Xiong, Kingsum Chow, Shuibing He, and Shuiguang Deng. Toward robust and efficient ML-based GPU caching for modern inference (lcr/laru).arXiv preprint arXiv:2509.20979, 2025. URLhttps://arxiv.org/abs/2509.20979

    Pith/arXiv arXiv 2025

  13. [13]

    Token economics for LLM agents

    Yuxi Chen, Junming Chen, Chenyu He, Yiwei Li, et al. Token economics for LLM agents. arXiv preprint arXiv:2605.09104, 2026. URLhttps://arxiv.org/abs/2605.09104

    Pith/arXiv arXiv 2026

  14. [14]

    Network offloading policies for 30 cloud robotics: A learning-based approach

    Sandeep Chinchali, Apoorva Sharma, James Harrison, Amine Elhafsi, Daniel Kang, Evgenya Pergament, Eyal Cidon, Sachin Katti, and Marco Pavone. Network offloading policies for 30 cloud robotics: A learning-based approach. InProceedings of Robotics: Science and Systems (RSS), Freiburg im Breisgau, Germany, 2019. doi: 10.15607/RSS.2019.XV.063. URL https://www...

    work page doi:10.15607/rss.2019.xv.063 2019

  15. [15]

    Server memory prices could double by 2026 as AI demand strains supply, November 2025

    Counterpoint Research. Server memory prices could double by 2026 as AI demand strains supply, November 2025. URLhttps://www.networkworld.com/article/4093752/serve r-memory-prices-could-double-by-2026-as-ai-demand-strains-supply.html . Via Network World; 32GB DDR5 module $149 to $239; server DDR5 $1.50/Gb

    arXiv 2026

  16. [16]

    Flashield: a hybrid key-value cache that controls flash write amplification

    Assaf Eisenman, Asaf Cidon, Evgenya Pergament, Or Haimovich, Ryan Stutsman, Mohammad Alizadeh, and Sachin Katti. Flashield: a hybrid key-value cache that controls flash write amplification. In16th USENIX Symposium on Networked Systems Design and Implementation (NSDI ’19), pages 65–78. USENIX Association, 2019. URLhttps://www.usenix.org/confe rence/nsdi19/...

    2019

  17. [17]

    NAND flash prices are surging in 2026: +33–38% q1, +70–75% q2, April 2026

    Elinfor (relaying TrendForce). NAND flash prices are surging in 2026: +33–38% q1, +70–75% q2, April 2026. URLhttps://www.elinfor.com/knowledge/nand-flash-prices-are-sur ging-in-2026-what-it-means-for-your-supply-chain-and-how-to-prepare-p-11312

    2026

  18. [18]

    B200 cost breakdown, 2024

    Epoch AI. B200 cost breakdown, 2024. URLhttps://epoch.ai/data-insights/b200-cos t-breakdown. HBM3E component cost $14–17/GB on 192 GB B200

    2024

  19. [19]

    Regulation (eu) 2024/1781 establishing a framework for the setting of ecodesign requirements for sustainable products (espr) and the digital product passport, 2024

    European Parliament and Council. Regulation (eu) 2024/1781 establishing a framework for the setting of ecodesign requirements for sustainable products (espr) and the digital product passport, 2024. URL https://data.europa.eu/en/news-events/news/eus-digital -product-passport-advancing-transparency-and-sustainability . In force 18 Jul 2024; product-specific...

    2024

  20. [20]

    Directive (eu) 2024/1799 on common rules promoting the repair of goods (right-to-repair directive), 2024

    European Parliament and Council. Directive (eu) 2024/1799 on common rules promoting the repair of goods (right-to-repair directive), 2024. URLhttps://commission.europa.eu/law/l aw-topic/consumer-protection-law/directive-repair-goods_en. Adopted 13 Jun 2024; in force 30 Jul 2024; transposition by 31 Jul 2026; servers and data-storage products in Annex II

    2024

  21. [21]

    John C. Gittins. Bandit processes and dynamic allocation indices.Journal of the Royal Statistical Society, Series B, 41(2):148–164, 1979. doi: 10.1111/j.2517-6161.1979.tb01068.x

    work page doi:10.1111/j.2517-6161.1979.tb01068.x 1979

  22. [22]

    Wise, Alberto Speranzon, and Luca Carlone

    Nicolas Gorlo, Derek K. Wise, Alberto Speranzon, and Luca Carlone. Worth remembering: Surprise-gated robot episodic memory, 2026. URLhttps://arxiv.org/abs/2606.03787

    Pith/arXiv arXiv 2026

  23. [23]

    Lee, David Brooks, and Carole-Jean Wu

    Udit Gupta, Mariam Elgamal, Gage Hills, Gu-Yeon Wei, Hsien-Hsin S. Lee, David Brooks, and Carole-Jean Wu. ACT: Designing sustainable computer systems with an architectural carbon modeling tool. InProc. 49th Annual Int. Symp. Computer Architecture (ISCA), 2022. doi: 10.1145/3470496.3527408. URLhttps://ugupta.com/files/Gupta_ISCA2022_ACT.pdf

    work page doi:10.1145/3470496.3527408 2022

  24. [24]

    Hall and Dale W

    Robert E. Hall and Dale W. Jorgenson. Tax policy and investment behavior.American Economic Review, 57(3):391–414, 1967. URLhttp://piketty.pse.ens.fr/files/HallJor genson67.pdf

    1967

  25. [25]

    The economics of exhaustible resources.Journal of Political Economy, 39(2): 137–175, 1931

    Harold Hotelling. The economics of exhaustible resources.Journal of Political Economy, 39(2): 137–175, 1931. doi: 10.1086/254195. 31

    work page doi:10.1086/254195 1931

  26. [26]

    Rotman, P

    Nathan Jay, Noga H. Rotman, P. Brighten Godfrey, Michael Schapira, and Aviv Tamar. A deep reinforcement learning perspective on internet congestion control. InProceedings of the 36th International Conference on Machine Learning (ICML), pages 3050–3059, 2019. URL http://proceedings.mlr.press/v97/jay19a/jay19a.pdf

    2019

  27. [27]

    Jorgenson

    Dale W. Jorgenson. The theory of investment behavior. In Robert Ferber, editor,Determinants of Investment Behavior, pages 129–175. National Bureau of Economic Research (Columbia Univ. Press), 1967. URLhttps://www.nber.org/system/files/chapters/c1235/c1235.pdf

    1967

  28. [28]

    Neurosurgeon: Collaborative intelligence between the cloud and mobile edge.ACM SIGARCH Computer Architecture News (ASPLOS ’17), 45(1):615–629, 2017

    Yiping Kang, Johann Hauswald, Cao Gao, Austin Rovinski, Trevor Mudge, Jason Mars, and Lingjia Tang. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge.ACM SIGARCH Computer Architecture News (ASPLOS ’17), 45(1):615–629, 2017. doi: 10.1145/3093337.3037698. URLhttps://dl.acm.org/doi/10.1145/3093337.3037698

    work page doi:10.1145/3093337.3037698 2017

  29. [29]

    DROID: A large-scale in-the-wild robot manipulation dataset, 2024

    Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, et al. DROID: A large-scale in-the-wild robot manipulation dataset, 2024. URL https://arxiv.org/abs/2403.12945

    Pith/arXiv arXiv 2024

  30. [30]

    Edge computing and GDPR: A technical security and legal compliance analysis

    Jonathan Kilit and Joel Bobin Blychert. Edge computing and GDPR: A technical security and legal compliance analysis. Bachelor’s thesis, Jönköping University, School of Engineering, 2025. URL https://www.diva-portal.org/smash/get/diva2:1982107/FULLTEXT01.pdf. GDPR Arts. 5, 25, 32, 44, 48; edge processing for data residency and the GDPR vs. US CLOUD Act conflict

    2025

  31. [31]

    OpenVLA: An open-source vision-language-action model, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. OpenVLA: An open-source vision-language-action model, 2024. URLhttps: //arxiv.org/abs/2...

    Pith/arXiv arXiv 2024

  32. [32]

    Understanding TBW versus P/E cycles in managed flash memory

    KIOXIA. Understanding TBW versus P/E cycles in managed flash memory. KIOXIA technical brief, 2023. URL https://americas.kioxia.com/content/dam/kioxia/en-us/business/m emory/mlc-nand/asset/KIOXIA-TBW-vs-PE-Cycles-Tech-Brief.pdf . P/E endurance for SLC/MLC/TLC; eMMC/UFS rated in P/E cycles

    2023

  33. [33]

    Managed-retention memory: A new class of memory for the ai era.arXiv preprint arXiv:2501.09605, 2025

    Sergey Legtchenko, Ioan Stefanovici, Richard Black, Antony Rowstron, Junyi Liu, Paolo Costa, Burcu Canakci, Dushyanth Narayanan, and Xingbo Wu. Managed-retention memory: A new class of memory for the ai era.arXiv preprint arXiv:2501.09605, 2025. URL https: //arxiv.org/abs/2501.09605

    arXiv 2025

  34. [34]

    LIBERO: Benchmarking knowledge transfer for lifelong robot learning, 2023

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: Benchmarking knowledge transfer for lifelong robot learning, 2023. URLhttps: //arxiv.org/abs/2306.03310

    Pith/arXiv arXiv 2023

  35. [35]

    Liu, Milad Hashemi, Kevin Swersky, Parthasarathy Ranganathan, and Junwhan Ahn

    Evan Z. Liu, Milad Hashemi, Kevin Swersky, Parthasarathy Ranganathan, and Junwhan Ahn. An imitation learning approach for cache replacement. InProceedings of the 37th International Conference on Machine Learning (ICML), pages 6237–6247, 2020. URLhttps: //proceedings.mlr.press/v119/liu20f.html

    2020

  36. [36]

    Proactive retention-aware online video caching scheme in mobile edge computing (dpro).Computer Communications, 239:108313, 2025

    Guangzhou Liu, Zhen Qian, and Guanghui Li. Proactive retention-aware online video caching scheme in mobile edge computing (dpro).Computer Communications, 239:108313, 2025. doi: 32 10.1016/j.comcom.2025.108313. URL https://www.sciencedirect.com/science/article/ abs/pii/S0140366425002701

    work page doi:10.1016/j.comcom.2025.108313 2025

  37. [37]

    Resource man- agement with deep reinforcement learning

    Hongzi Mao, Mohammad Alizadeh, Ishai Menache, and Srikanth Kandula. Resource man- agement with deep reinforcement learning. InProceedings of the 15th ACM Workshop on Hot Topics in Networks (HotNets), pages 50–56, 2016. doi: 10.1145/3005745.3005750. URL https://people.csail.mit.edu/alizadeh/papers/deeprm-hotnets16.pdf

    work page doi:10.1145/3005745.3005750 2016

  38. [38]

    Split computing and early exiting for deep learning applications: Survey and research challenges.ACM Computing Surveys, 55(5):1–30, 2022

    Yoshitomo Matsubara, Marco Levorato, and Francesco Restuccia. Split computing and early exiting for deep learning applications: Survey and research challenges.ACM Computing Surveys, 55(5):1–30, 2022. doi: 10.1145/3527155. URLhttps://arxiv.org/abs/2103.04505

    work page doi:10.1145/3527155 2022

  39. [39]

    Berger, Nathan Beckmann, and Gregory R

    Sara McAllister, Benjamin Berg, Julian Tutuncu-Macias, Juncheng Yang, Sathya Gunasekar, Jimmy Lu, Daniel S. Berger, Nathan Beckmann, and Gregory R. Ganger. Kangaroo: Caching billions of tiny objects on flash. InProceedings of the ACM SIGOPS 28th Symposium on Operating Systems Principles (SOSP ’21), pages 243–262, 2021. doi: 10.1145/3477132.3483568. URLhtt...

    work page doi:10.1145/3477132.3483568 2021

  40. [40]

    Le, Benoit Steiner, Rasmus Larsen, Yuefeng Zhou, Naveen Kumar, Mohammad Norouzi, Samy Bengio, and Jeff Dean

    Azalia Mirhoseini, Hieu Pham, Quoc V. Le, Benoit Steiner, Rasmus Larsen, Yuefeng Zhou, Naveen Kumar, Mohammad Norouzi, Samy Bengio, and Jeff Dean. Device placement opti- mization with reinforcement learning. InProceedings of the 34th International Conference on Machine Learning (ICML), pages 2430–2439, 2017. URLhttps://proceedings.mlr.press/ v70/mirhosein...

    2017

  41. [41]

    SSD endurance and NAND types explained for 2026: TLC, QLC and more

    Newegg Insider. SSD endurance and NAND types explained for 2026: TLC, QLC and more. Newegg Insider, 2026. URL https://www.newegg.com/insider/ssd-lifespan-decod ed-understanding-nand-types-and-write-endurance-in-2026/ . Quality TLC NVMe: 1,500–3,000 P/E; 1 TB TLC = 300–1,200 TBW

    2026

  42. [42]

    NVIDIA Jetson AGX Orin series technical brief

    NVIDIA. NVIDIA Jetson AGX Orin series technical brief. NVIDIA technical brief, 2022. URL https://www.nvidia.com/content/dam/en-zz/Solutions/gtcf21/jetson-orin/nvidia -jetson-agx-orin-technical-brief.pdf. 64 GB 256-bit LPDDR5, 204.8 GB/s; 15–60 W; up to 275 TOPS INT8

    2022

  43. [43]

    Introducing NVIDIA Jetson Thor, the ultimate platform for physical AI

    NVIDIA. Introducing NVIDIA Jetson Thor, the ultimate platform for physical AI. NVIDIA Developer Blog, 2025. URLhttps://developer.nvidia.com/blog/introducing-nvidia-j etson-thor-the-ultimate-platform-for-physical-ai/ . Jetson AGX Thor / T5000: 128 GB 256-bit LPDDR5X, 273 GB/s, 40–130 W, up to 2070 FP4 TFLOPS

    2025

  44. [44]

    Agent memory: Characterization and system implications of stateful long-horizon workloads.arXiv preprint arXiv:2606.06448, 2026

    Yasmine Omri, Ziyu Gan, Zachary Broveak, Robin Geens, Zexue He, Alex Pentland, Marian Verhelst, Tsachy Weissman, and Thierry Tambe. Agent memory: Characterization and system implications of stateful long-horizon workloads.arXiv preprint arXiv:2606.06448, 2026. URL https://arxiv.org/abs/2606.06448

    Pith/arXiv arXiv 2026

  45. [45]

    Patil, Ion Stoica, and Joseph E

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. MemGPT: Towards LLMs as operating systems, 2023. URLhttps: //arxiv.org/abs/2310.08560

    Pith/arXiv arXiv 2023

  46. [46]

    Beyond per-token pricing: A concurrency-aware methodology for LLM infras- tructure cost estimation, 2026

    Chitral Patil. Beyond per-token pricing: A concurrency-aware methodology for LLM infras- tructure cost estimation, 2026. URLhttps://arxiv.org/abs/2606.11690. Underutilization penalty 2.5–24x (1–10 rps), up to 36.3x near idle; cost off by exactly 1/U. 33

    Pith/arXiv arXiv 2026

  47. [47]

    Assessing the embodied carbon footprint of IoT edge devices with a bottom-up life-cycle approach.Journal of Cleaner Production, 2021

    Thibault Pirson and David Bol. Assessing the embodied carbon footprint of IoT edge devices with a bottom-up life-cycle approach.Journal of Cleaner Production, 2021. doi: 10.1016/j.jcle pro.2021.128966. URLhttps://arxiv.org/abs/2105.02082

    work page doi:10.1016/j.jcle 2021

  48. [48]

    What is QLC SSD

    Pure Storage. What is QLC SSD. Pure Storage knowledge base, 2025. URLhttps://www. everpuredata.com/knowledge/what-is-qlc-flash.html . QLC 1,000 P/E cycles; SLC 100,000 P/E cycles

    2025

  49. [49]

    FrozenHot cache: Rethinking cache management for modern hardware

    Ziyue Qiao, Xiaocheng Wu, Yiding Zhang, Yang Gao, Yuhao Zhou, Juncheng Yang, et al. FrozenHot cache: Rethinking cache management for modern hardware. InProceedings of the 18th European Conference on Computer Systems (EuroSys), 2023. doi: 10.1145/3552326.3587446. URLhttps://www.pdl.cmu.edu/ftp/Storage/FrozenHot-Eurosys23.pdf

    work page doi:10.1145/3552326.3587446 2023

  50. [51]

    URLhttps://arxiv.org/abs/2401.00448

    arXiv

  51. [52]

    Life-cycle emissions of AI hardware: A cradle-to-grave approach and generational trends, 2025

    Ian Schneider, Hui Xu, Stephan Benecke, David Patterson, Keguo Huang, Parthasarathy Ranganathan, and Cooper Elsworth. Life-cycle emissions of AI hardware: A cradle-to-grave approach and generational trends, 2025. URLhttps://arxiv.org/abs/2502.01671

    arXiv 2025

  52. [53]

    SmolVLA: A vision-language-action model for affordable and efficient robotics, 2025

    Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, Simon Alibert, Matthieu Cord, Thomas Wolf, and Remi Cadene. SmolVLA: A vision-language-action model for affordable and efficient robotics, 2025. URLhttps://arxiv.org/abs/2506.01844

    Pith/arXiv arXiv 2025

  53. [54]

    Berger, Kai Li, and Wyatt Lloyd

    Zhenyu Song, Daniel S. Berger, Kai Li, and Wyatt Lloyd. Learning relaxed Belady for content distribution network caching. In17th USENIX Symposium on Networked Systems Design and Implementation (NSDI), pages 529–544, 2020. URLhttps://www.usenix.org/system/file s/nsdi20-paper-song.pdf

    2020

  54. [55]

    MemER: Scaling up memory for robot control via experience retrieval, 2025

    Ajay Sridhar, Jennifer Pan, Satvik Sharma, and Chelsea Finn. MemER: Scaling up memory for robot control via experience retrieval, 2025. URLhttps://arxiv.org/abs/2510.20328

    arXiv 2025

  55. [56]

    Junkyard computing: Repurposing discarded smartphones to minimize carbon

    Jennifer Switzer, Gabriel Marcano, Ryan Kastner, and Pat Pannuto. Junkyard computing: Repurposing discarded smartphones to minimize carbon. InProc. 28th ACM Int. Conf. Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2023. doi: 10.48550/arXiv.2110.06870. URLhttps://arxiv.org/abs/2110.06870

    work page doi:10.48550/arxiv.2110.06870 2023

  56. [57]

    Higher DDR5 profitability intensifies capacity crowding; HBM3e–DDR5 asp gap to narrow from 4–5x to 1–2x by end-2026, December 2025

    TrendForce. Higher DDR5 profitability intensifies capacity crowding; HBM3e–DDR5 asp gap to narrow from 4–5x to 1–2x by end-2026, December 2025. URLhttps://www.trendforce.c om/presscenter/news/20251218-12843.html

    arXiv 2026

  57. [58]

    Efficient ssd caching by avoiding unnecessary writes using machine learning

    Hua Wang, Xinbo Yi, Ping Huang, Bin Cheng, and Ke Zhou. Efficient ssd caching by avoiding unnecessary writes using machine learning. InProceedings of the 47th International Conference on Parallel Processing / ACM, 2018. doi: 10.1145/3225058.3225126. URL https://dl.acm.org/doi/10.1145/3225058.3225126

    work page doi:10.1145/3225058.3225126 2018

  58. [59]

    KARMA: Augmenting embodied AI agents with long-and-short term memory systems, 2024

    Zixuan Wang, Bo Yu, Junzhe Zhao, Wenhao Sun, Sai Hou, Shuai Liang, Xing Hu, Yinhe Han, and Yiming Gan. KARMA: Augmenting embodied AI agents with long-and-short term memory systems, 2024. URLhttps://arxiv.org/abs/2409.14908. 34

    arXiv 2024

  59. [60]

    Embodied carbon footprint of 3D NAND memories

    Olivier Weppe, Thibaut Marty, Sylvain Toussaint, Nicolas Brusselmans, Jean-Christophe Prévotet, Jean-Pierre Raskin, and Maxime Pelcat. Embodied carbon footprint of 3D NAND memories. InProceedings of the 22nd ACM International Conference on Computing Frontiers: Workshops and Special Sessions (CF ’25 Companion), 2025. doi: 10.1145/3706594.3727962. URL https...

    work page doi:10.1145/3706594.3727962 2025

  60. [61]

    Western Digital industrial flash storage portfolio (industrial e.mmc/ufs/ssd)

    Western Digital. Western Digital industrial flash storage portfolio (industrial e.mmc/ufs/ssd). Product portfolio brochure, 2023. URLhttps://www.marubun.co.jp/wp-content/uploads /a7ijkd000000dgaz/a7ijkd000000m6x1.pdf. Industrial 3D-NAND: 3K P/E high-endurance grade; up to 1,600 TBW

    2023

  61. [62]

    Restless bandits: Activity allocation in a changing world.Journal of Applied Probability, 25A:287–298, 1988

    Peter Whittle. Restless bandits: Activity allocation in a changing world.Journal of Applied Probability, 25A:287–298, 1988. doi: 10.2307/3214163

    work page doi:10.2307/3214163 1988

  62. [63]

    Berger, Nathan Beckmann, and Gregory R

    Daniel Lin-Kit Wong, Hao Wu, Carson Molder, Sathya Gunasekar, Jimmy Lu, Snehal Khandkar, Abhinav Sharma, Daniel S. Berger, Nathan Beckmann, and Gregory R. Ganger. Baleen: ML admission & prefetching for flash caches. In22nd USENIX Conference on File and Storage Technologies (FAST), 2024. URLhttps://www.usenix.org/conference/fast24/presentat ion/wong

    2024

  63. [64]

    Cachesack: Theory and experience of google’s admission optimization for datacenter flash caches.ACM Transactions on Storage, 19(2):1–24, 2023

    Tzu-Wei Yang, Seth Pollen, Mustafa Uysal, Arif Merchant, Homer Wolfmeister, and Junaid Khalid. Cachesack: Theory and experience of google’s admission optimization for datacenter flash caches.ACM Transactions on Storage, 19(2):1–24, 2023. doi: 10.1145/3582014. URL https://dl.acm.org/doi/10.1145/3582014

    work page doi:10.1145/3582014 2023

  64. [65]

    Nemo: A low-write-amplification cache for tiny objects on log-structured flash devices

    Xufeng Yang, Tingting Tan, Jing Hu, Congming Gao, Mingyan Liu, Tianyang Jiang, Jian Chen, Linbo Long, Yina Lv, and Jiwu Shu. Nemo: A low-write-amplification cache for tiny objects on log-structured flash devices. InProceedings of the 31st ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS ’26), Vo...

    work page doi:10.1145/3779212.3790191 2026

  65. [66]

    A machine learning based write policy for ssd cache in cloud block storage

    Ji Zhang, Ke Zhou, et al. A machine learning based write policy for ssd cache in cloud block storage. InDesign, Automation & Test in Europe Conference (DATE), pages 82–87, 2020. URLhttps://past.date-conference.com/proceedings-archive/2020/pdf/0022.pdf

    2020

  66. [67]

    Agentic AI systems should be designed as marginal token allocators.arXiv preprint arXiv:2605.01214, 2026

    Siqi Zhu. Agentic AI systems should be designed as marginal token allocators.arXiv preprint arXiv:2605.01214, 2026. URLhttps://arxiv.org/abs/2605.01214. 35

    Pith/arXiv arXiv 2026