Pith. sign in

REVIEW 4 major objections 4 minor 79 references

Caching for the Future: Scrub Jay Episodic Memory Principles for Agent Memory Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Agent memories should not all age at the same rate; per-memory, type-conditioned decay is what lets a retrieval system generalize to retention intervals it never calibrated on.

desk verdict Honest, reproducible architecture paper whose central mechanism claim is undercut by a benchmark that shares its taxonomy with the system. read the letter →

arxiv 2608.04746 v1 pith:52Z5TDGO submitted 2026-08-05 cs.CL cs.IR

classification cs.CLcs.IR
keywords episodicmemoryLLMagentsdecayperishabilitytemporalgeneralizationWhat-Where-WhenretrievalGapagent
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

LLM agents that remember across sessions currently treat all memories as equally persistent, so retrievals slowly fill with outdated facts. This paper argues that the fix is to give each memory its own perishability coefficient, a per-memory, type-conditioned decay rate inspired by how scrub jays treat worms and peanuts differently as caches age. It builds ScrubJay-MEM, an external memory store in which every memory is bound into a what-where-when tuple with an auto-estimated decay rate and horizon, and introduces the Temporal Generalization Test (TGT) to ask whether a system can generalize to retention intervals it never calibrated on. On TGT, ScrubJay-MEM is the only retrieval-based system with clearly positive Generalization Gap (+0.108), and removing the decay collapses the gap 5.7-fold while leaving staleness accuracy unchanged. A sympathetic reader should care because this isolates an architectural mechanism, temporal validity as an inspectable per-memory parameter, rather than just another retrieval score.

What carries the argument

The central object is the perishability coefficient $\pi_i$ and its companion utility horizon $\tau_i$, stored with every memory as part of a jointly-bound What–Where–When tuple. The utility function $U(m_i,t_q)=V_i\exp(-\pi_i(t_q-t_i)/\tau_i)$ is what changes how the system treats memory as it ages: stable knowledge has $\pi\approx0.05$ and persists for months, while ephemeral facts have $\pi\approx0.9$ and expire within hours. A four-factor retrieval score with query-adaptive weights combines what, where, when, and graph-neighborhood signals; a Retroactive Contextual Integration pass revises $V$, $\pi$, and $\tau$ in parameter space at O(1) LLM calls per update; and the Temporal Generalization Test measures whether a system's accuracy at an unseen middle retention interval matches linear interpolation from the flanking seen intervals. The mechanism that carries the argument is the decay term: removing it collapses GenGap by $5.7\times$.

What would settle it

Build a second TGT whose ground-truth validity schedule comes from an independent source, such as human annotations of when sample facts expire or a different taxonomy with different perishability classes, and run ScrubJay-MEM unchanged; if GenGap falls toward zero or negative, the claimed generalization was an artifact of the shared taxonomy rather than a property of type-conditioned decay.

Watch

Extended reading notes

Core claim

The central claim is that temporal generalization in LLM-agent memory requires per-memory, type-conditioned decay, and that such decay can be computed cheaply and made inspectable. Each memory is stored as an Episodic Memory Unit, a jointly bound What–Where–When tuple with a perishability coefficient $\pi_i$ and utility horizon $\tau_i$; retrieval scores combine semantic similarity, context similarity, a perishability-aware utility $U(m_i,t_q)=V_i\exp(-\pi_i(t_q-t_i)/\tau_i)$, and a graph bonus through query-adaptive weights. Evaluated on the new Temporal Generalization Test, ScrubJay-MEM achieves GenGap $+0.108$, the only retrieval-only system above zero, and the no-decay ablation reduces GenGap by $5.7\times$ while staleness accuracy stays at 82.3%, which the authors present as establishing decay as the responsible mechanism. On MemoryAgentBench EventQA-64k with an 8B backbone, it reaches 61.58 F1, +2.66 over Mem0 and +3.09 over Qwen3-Embedding-4B; with a stronger 30B backbone the gap narrows, and on fact-consolidation tasks the decay is actively miscalibrated, scoping the contribution to temporal reasoning over perishable facts.

Load-bearing premise

The benchmark's ground truth says which memories are stale at each retention interval, and that validity schedule was built from the same four perishability classes and decay ranges that ScrubJay-MEM's own classifier uses, so the positive generalization result could partly reflect the system and the test sharing the same view of what decays.

Editorial extensions

If this is right

  • A memory store can expose temporal validity as an inspectable parameter, so an agent can know why a memory was retrieved, not just what ranked high, and can suppress stale items before they enter context.
  • Uniform recency decay is not enough: on TGT a global recency prior scores below a flat dense retriever, so decay must be conditioned on memory type.
  • Retroactive updates can revise many memories through one LLM call plus vector operations, keeping maintenance cost independent of store size.
  • Gains are concentrated in the small-backbone regime; with a 30B LLM, dense retrieval matches or beats the architecture, so the decay mechanism is complementary to model capacity rather than a replacement for it.
  • The approach is scoped to perishable-fact reasoning; on fact-consolidation tasks, where stale facts must remain visible, the same decay is miscalibrated and flat retrieval is preferable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A decisive follow-up test would build TGT's validity schedule from an independent source, such as human expiration judgments or a different taxonomy, so that the system's classifier and the benchmark's ground truth do not share categorical primitives; the paper's own limitations note this shared-taxonomy risk.
  • If per-memory decay proves to be the right primitive, the same $(\pi_i,\tau_i)$ pair could serve as a general time-to-live signal outside LLM agents, for example in cache invalidation or for suppressing stale documents in any retrieval stack.
  • The auto-classification step is currently an LLM prompt with a keyword fallback; replacing it with a learned predictor trained on user-confirmed staleness would make perishability personal and could enlarge the GenGap margin.
  • Because gains narrow under stronger backbones, the architecture's natural deployment is small, locally served LLMs; one could combine it with a strong LLM that handles consolidation while decay handles retrieval, rather than treating the two as competitors.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes ScrubJay-MEM, an LLM-agent memory store in which each memory is encoded as a What-Where-When tuple with an auto-classified perishability coefficient π_i and utility horizon τ_i, retrieved by query-adaptive scoring that combines semantic, contextual, temporal-utility, and graph-bonus terms, and maintained by retroactive parameter updates and a prospective buffer. The authors evaluate on MemoryAgentBench EventQA-64k and introduce the Temporal Generalization Test (TGT), a synthetic benchmark with held-out retention intervals and a GenGap metric; they report that ScrubJay-MEM is the only retrieval-based system with substantially positive GenGap (+0.108) and that removing decay collapses GenGap by 5.7×, which they interpret as establishing type-conditioned decay as necessary for temporal generalization. The paper also reports honest scoping results: gains on EventQA shrink or reverse under a stronger backbone, and the system underperforms flat retrieval on fact-consolidation subsets.

Significance. If the central GenGap claim were valid, ScrubJay-MEM would offer a valuable, inspectable architectural signal for temporal validity in agent memory, at no per-query LLM cost, and the TGT benchmark could be a useful diagnostic. The paper has genuine strengths: it provides formal boundedness/contraction proofs for the retroactive update, a detailed benchmark-construction appendix with explicit quality audits and leakage-prevention checks, full hyperparameter disclosure, and unusually candid limitations and failure analyses. These strengths, however, do not rescue the central mechanistic claim, because the TGT validity schedule is constructed from the same four-class perishability taxonomy that the system's own classifier uses, and because the GenGap metric as defined rewards systems with no temporal modeling at all. The comparative contributions on EventQA-64k are modest, backbone-dependent, and reverse under the stronger backbone, so the paper's main positive result rests on the circular benchmark.

major comments (4)
  1. [§4.2, Tables 5 and 6, Appendix F.3, Limitations]
  2. [§4.2, Eq. (16), Table 2, Figure 3]
  3. [§4.2, Tables 2 and 8]
  4. [§4.1, Table 1, Appendix C, Table 11]
minor comments (4)
  1. [Throughout]
  2. [References]
  3. [§3.6, Proposition 2]
  4. [§4.2 and Appendix B.3]

Circularity Check

1 steps flagged · score 8.0 of 10

TGT's validity schedule is built from the same four-class taxonomy and π-ranges as ScrubJay-MEM's own classifier, so the +0.108 GenGap and its decay-necessity reading are partly forced by construction.

  1. self definitional [Limitations; TGT construction (Appendix B.1/B.7, Table 5); perishability taxonomy (Table 6); Eq. 3 utility; Appendix F.3 classifier fallback]
    "TGT is a controlled diagnostic. TGT is LLM-naturalised from deterministic templates with held-out retention intervals and a known validity schedule, properties that enable a mechanistic claim about generalisation but limit external validity to in-distribution conversational logs. TGT’s four perishability classes share π-ranges with our internal taxonomy (Tables 3 and 6); the no-decay ablation, which collapses GenGap by 5.7×, controls for the concern that the benchmark and the architecture share categorical primitives."

    Concrete reduction: TGT correctness per query is the type×interval schedule of Table 5, and Table 6 defines those same four types with the π/τ ranges ScrubJay-MEM's taxonomy uses. The F.3 fallback assigns π=0.9/0.6/0.3/0.1 to ephemeral/task-specific/procedural/factual, the midpoints of A/B/C/D ranges. Plugged into U=Vi exp(−π Δt/τ), these suppress exactly the cells Table 5 marks invalid and preserve exactly those it marks valid, so GenGap measures classifier–benchmark agreement. The no-decay ablation removes only the decay term while the benchmark remains built from the same taxonomy, so it does not control for the acknowledged sharing; the necessity claim is forced by construction.

full rationale

The TGT benchmark's ground-truth validity schedule (Table 5) is defined by four perishability classes, and those same four classes with the same π and τ ranges are what ScrubJay-MEM's own classifier encodes (Table 6, Appendix F.3). The system's decay utility (Eq. 3) therefore reproduces the benchmark's correctness labels by construction, making the reported +0.108 GenGap a measure of agreement between the classifier's internal taxonomy and the benchmark's validity schedule rather than an independent test of temporal generalization. The paper honestly acknowledges this sharing, but the no-decay ablation that it offers as a control does not break the alignment: the ablation removes only the decay term, not the shared categorical primitives that make the decay curve match the schedule. Consequently the central claim that type-conditioned decay is necessary for the TGT result reduces to a benchmark-construction choice. The EventQA-64k result is an external benchmark and is not part of this circular step, which is why the score is 8 rather than 10; the circularity is load-bearing on the paper's mechanism-necessity claim but not on every empirical result in the paper.

Assumptions & free parameters 12 free parameters · 6 assumptions · 0 invented entities

The TGT result relies on a benchmark whose validity schedule is created from the same perishability taxonomy the system uses, and on a large set of hand-set hyperparameters. The only mathematically guaranteed properties (Proposition 1) are elementary boundedness and contraction facts that do not validate the empirical claims.

free parameters (12)
  • WWW binding fusion weights = 0.55, 0.35, 0.10
    Fixed convex mixture in Eq. 2 for the EMU binding vector; hand-chosen, no learning.
  • Graph edge mixture weights (wS, wC, wT) = 0.40, 0.30, 0.30
    Eq. 7 combines semantic, contextual, temporal edges with hand-set weights.
  • Retrieval score base logits and keyword biases = 1.5, 1.0, 1.0, 1.0 plus keyword increments
    Eq. 6 adaptive weights are produced by a hand-designed heuristic scoring function.
  • RCI learning rates (eta_V, eta_pi, eta_tau, eta_w) = 0.20, 0.15, 0.15, 0.20
    Appendix F, Table 13; control update magnitudes in Eqs. 9-12.
  • theta_RCI = 0.60
    Cosine gate for affected-memory selection in RCI (Section 3.5).
  • Perishability pi ranges per memory type = factual 0.05-0.15, procedural 0.2-0.4, task-specific 0.5-0.7, ephemeral 0.8-1.0
    Tables 3 and 6; the core type-conditioning, hand-assigned, and identical to TGT ground-truth ranges.
  • Utility horizon tau ranges per type = weeks-months, days-weeks, hours-days, minutes-hours
    Tables 3 and 6; hand-assigned horizon bands.
  • Keyword-fallback classifier values = ephemeral pi=0.9 tau=2h, procedural pi=0.3 tau=10d, task pi=0.6 tau=24h, factual pi=0.1 tau=45d
    Appendix F.3; deterministic fallback when LLM parser fails.
  • tau_min, tau_max = 60 s, 90 days
    Utility-horizon clamp in Eq. 11 and Table 13.
  • Pruning thresholds epsilon, epsilon_G = 0.05, 0.05
    Eq. 15 retention rule; hand-set.
  • PMB confidence threshold theta_B = 0.45
    Buffer early-exit threshold in Section 3.6.
  • Temporal edge decay lambda_T and window Delta_T = 1/86400 s^-1, 7 days
    Eq. 7 temporal edge kernel; hand-set.
assumptions (6)
  • domain assumption Exponential utility decay U(m_i,t) = V_i exp(-pi_i (t-t_i)/tau_i) models memory validity
    Postulated in Eq. 3 without empirical calibration; central to the mechanism.
  • ad hoc to paper Four perishability classes span the realistic decay space
    The taxonomy in Tables 3 and 6 is introduced by the authors and also used as TGT ground truth.
  • ad hoc to paper TGT ground-truth validity schedule (Table 5) is correct and independent of the system
    The benchmark is synthetic and uses the same class structure as the system, so this independence fails.
  • domain assumption LLM prompt classifier estimates pi and tau without training
    No calibration or accuracy analysis of the classifier is provided.
  • domain assumption Query-adaptive heuristic weights improve over fixed weights
    No ablation compares Eq. 6 to fixed weights.
  • standard math Banach fixed-point theorem and non-expansiveness of L2 normalization
    Used in Proposition 1 proof, Appendix E.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Caching for the Future: Scrub Jay Episodic Memory Principles for Agent Memory Systems." pith.science (2026). https://pith.science/paper/52Z5TDGO

@misc{pith2026260804746,
  author       = {Pith},
  title        = {Pith review of: Caching for the Future: Scrub Jay Episodic Memory Principles for Agent Memory Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/52Z5TDGO}},
  note         = {Machine review of arXiv:2608.04746}
}
abstract

LLM agents that persist across sessions accumulate stored memories whose validity varies enormously by content type, yet existing memory architectures treat all memories as equally persistent and systematically contaminate retrieved context with outdated facts. We show that per-memory, type-conditioned temporal decay, a property of western scrub jay episodic memory, can be operationalized as an auto-classified coefficient $\pi_i$ in an external LLM-agent memory store, yielding ScrubJay-MEM: each memory is encoded as a jointly-bound What--Where--When tuple with an estimated perishability $\pi_i$ and utility horizon $\tau_i$, retrieved by query-adaptive scoring, and revised retroactively at $O(1)$ LLM calls per update. We introduce the Temporal Generalization Test (TGT), a benchmark with held-out retention intervals and a Generalization Gap (GenGap) metric. On TGT, ScrubJay-MEM is the only retrieval-based system with substantially positive GenGap ($+0.108$); on MemoryAgentBench EventQA-64k it improves F1 by $+2.66$ over Mem0 and $+3.09$ over Qwen3-Embedding-4B under a llm backbone. A decay ablation collapses GenGap by $5.7\times$, establishing type-conditioned decay as necessary for the result. Gains narrow under stronger backbones and reverse on fact-consolidation tasks, scoping the contribution to temporal reasoning over perishable facts.

Figures

Figures reproduced from arXiv: 2608.04746 by the authors.

Figure 1
Figure 1. ScrubJay-MEM architecture (blueprint view). [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Accuracy across the five retention intervals on E3 TGT. Shaded bands mark the two unseen intervals ( [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Generalization Gap across all evaluated systems on E3 TGT, sorted by GenGap. Among retrieval-based [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Operating-point view of E3 TGT: staleness accuracy vs. combined (factual [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Per-cell accuracy by memory type (A–D) and retention interval (I1–I5). Dashed columns mark the two held-out intervals. ScrubJay-MEM (full, left) outperforms the no-decay ablation (middle) most strongly on Type-A ephemeral memories at long intervals, in particular at I4…
Figure 6
Figure 6. Figure 6: Temporal Generalization Score (TGS) leader [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 49 canonical work pages

  1. [1]

    John R Anderson, Daniel Bothell, Michael D Byrne, Scott Douglass, Christian Lebiere, and Yulin Qin. 2004. An integrated theory of the mind. Psychological review, 111(4):1036

  2. [2]

    Cyril Brom and J Lukavsky. 2009. Towards virtual characters with a full episodic memory ii: The episodic memory strikes back. In Proc. empathic agents, AAMAS workshop, pages 1--9

  3. [3]

    Gordon DA Brown, Ian Neath, and Nick Chater. 2007. A temporal ratio model of memory. Psychological review, 114(3):539

  4. [5]

    Nicola S Clayton, Timothy J Bussey, and Anthony Dickinson. 2003. Can animals recall the past and plan for the future? Nature Reviews Neuroscience, 4(8):685--691

  5. [6]

    Nicola S Clayton and Anthony Dickinson. 1998. Episodic-like memory during cache recovery by scrub jays. Nature, 395(6699):272--274

  6. [7]

    Nicola S Clayton, Kara Shirley Yu, and Anthony Dickinson. 2001. Scrub jays (aphelocoma coerulescens) form integrated memories of the multiple features of caching episodes. Journal of Experimental Psychology: Animal Behavior Processes, 27(1):17

  7. [8]

    S \'e rgio PC Correia, Anthony Dickinson, and Nicola S Clayton. 2007. Western scrub-jays anticipate future needs independently of their current motivational state. Current Biology, 17(10):856--861

  8. [9]

    Jonathon D Crystal. 2010. Episodic-like memory in animals. Behavioural Brain Research, 215(2):235--243

Show all 79 references
  1. [10]

    Hermann Ebbinghaus. 1913. Memory

  2. [12]

    Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, and Jianfeng Gao. 2024. Model tells you what to discard: Adaptive kv cache compression for llms. In International Conference on Learning Representations, volume 2024, pages 22975--22988

  3. [13]

    Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Danihelka, Agnieszka Grabska-Barwi \'n ska, Sergio G \'o mez Colmenarejo, Edward Grefenstette, Tiago Ramalho, John Agapiou, and 1 others. 2016. Hybrid computing using a neural network with dynamic external memory. Natu...

  4. [16]

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised dense information retrieval with contrastive learning. Transactions on Machine Learning Research

  5. [19]

    Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. 2024. A human-inspired reading agent with gist memory of very long contexts. In International Conference on Machine Learning, pages 26396--26415. PMLR

  6. [20]

    Yaniv Leviathan, Matan Kalman, and Yossi Matias. 2023. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning, pages 19274--19286. PMLR

  7. [21]

    Zhixuan Lin, Evgenii Nikishin, Xu He, and Aaron Courville. 2025. Forgetting transformer: Softmax attention with a forget gate. In International Conference on Learning Representations, volume 2025, pages 69704--69738

  8. [23]

    Jesse Mu, Xiang Li, and Noah Goodman. 2023. Learning to compress prompts with gist tokens. Advances in Neural Information Processing Systems, 36:19327--19352

  9. [24]

    Charles Packer, Vivian Fang, Shishir\_G Patil, Kevin Lin, Sarah Wooders, and Joseph\_E Gonzalez. 2023. Memgpt: towards llms as operating systems

  10. [25]

    Arjun Panickssery, Samuel R Bowman, and Shi Feng. 2024. Llm evaluators recognize and favor their own generations. Advances in Neural Information Processing Systems, 37:68772--68802

  11. [26]

    Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1--22

  12. [27]

    Alexander Pritzel, Benigno Uria, Sriram Srinivasan, Adria Puigdomenech Badia, Oriol Vinyals, Demis Hassabis, Daan Wierstra, and Charles Blundell. 2017. Neural episodic control. In International conference on machine learning, pages 2827--2836. PMLR

  13. [28]

    Caroline R Raby, Dean M Alexis, Anthony Dickinson, and Nicola S Clayton. 2007. Planning for the future by western scrub-jays. Nature, 445(7130):919--921

  14. [31]

    Samuel Ritter, Ryan Faulkner, Laurent Sartran, Adam Santoro, Matthew Botvinick, and David Raposo. 2021. Rapid task-solving in novel environments. In International Conference on Learning Representations

  15. [32]

    Lucie H Salwiczek, Arii Watanabe, and Nicola S Clayton. 2010. Ten years of research into avian models of episodic-like memory and its implications for developmental and comparative cognition. Behavioural Brain Research, 215(2):221--234

  16. [33]

    Daniel L Schacter, Donna Rose Addis, and Randy L Buckner. 2007. Remembering the past to imagine the future: the prospective brain. Nature reviews neuroscience, 8(9):657--661

  17. [34]

    Alan Jay Smith. 1982. Cache memories. ACM Computing Surveys (CSUR), 14(3):473--530

  18. [35]

    Dennis Stachowicz and Geert-Jan M Kruijff. 2011. Episodic-like memory for cognitive robots. IEEE Transactions on Autonomous Mental Development, 4(1):1--16

  19. [36]

    Karl K Szpunar, R Nathan Spreng, and Daniel L Schacter. 2014. A taxonomy of prospection: Introducing an organizational framework for future-oriented cognition. Proceedings of the National Academy of Sciences, 111(52):18414--18421

  20. [37]

    Endel Tulving. 1972. Episodic and semantic memory. Organization of memory, 1(381-403):1

  21. [38]

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2025. Voyager: An open-ended embodied agent with large language models. Transactions on Machine Learning Research

  22. [39]

    Lei Wei, Xiao Peng, Xu Dong, Niantao Xie, and Bin Wang. 2026. https://arxiv.org/abs/2601.18642 Fademem: Biologically-inspired forgetting for efficient agent memory . Preprint, arXiv:2601.18642

  23. [40]

    Ella Worsfold, Nicola S Clayton, and Lucy G Cheke. 2025. Revisiting episodic-like memory in scrub jays: Is there more we can still learn from what--where--when caching behaviour? Learning & Behavior, 53(1):65--79

  24. [41]

    Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. 2025. Longmemeval: Benchmarking chat assistants on long-term interactive memory. In The Thirteenth International Conference on Learning Representations

  25. [42]

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2026. A-mem: Agentic memory for llm agents. Advances in Neural Information Processing Systems, 38:17577--17604

  26. [44]

    Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher R \'e , Clark Barrett, and 1 others. 2023. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information...

  27. [45]

    Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Memorybank: Enhancing large language models with long-term memory. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 19724--19731

  28. [46]

    Nature , volume=

    Episodic-like memory during cache recovery by scrub jays , author=. Nature , volume=. 1998 , publisher=

  29. [47]

    , author=

    Scrub jays (Aphelocoma coerulescens) form integrated memories of the multiple features of caching episodes. , author=. Journal of Experimental Psychology: Animal Behavior Processes , volume=. 2001 , publisher=

  30. [48]

    Nature Reviews Neuroscience , volume=

    Can animals recall the past and plan for the future? , author=. Nature Reviews Neuroscience , volume=. 2003 , publisher=

  31. [49]

    Nature , volume=

    Planning for the future by western scrub-jays , author=. Nature , volume=. 2007 , publisher=

  32. [50]

    Current Biology , volume=

    Western scrub-jays anticipate future needs independently of their current motivational state , author=. Current Biology , volume=. 2007 , publisher=

  33. [51]

    Advances in Neural Information Processing Systems , volume=

    A-mem: Agentic memory for llm agents , author=. Advances in Neural Information Processing Systems , volume=

  34. [52]

    , author=

    MemGPT: towards LLMs as operating systems. , author=. 2023 , publisher=

  35. [53]

    arXiv preprint arXiv:2504.19413 , year=

    Mem0: Building production-ready ai agents with scalable long-term memory , author=. arXiv preprint arXiv:2504.19413 , year=

  36. [54]

    arXiv preprint arXiv:2511.01448 , year=

    Licomemory: Lightweight and cognitive agentic memory for efficient long-term reasoning , author=. arXiv preprint arXiv:2511.01448 , year=

  37. [55]

    arXiv preprint arXiv:2501.13956 , year=

    Zep: a temporal knowledge graph architecture for agent memory , author=. arXiv preprint arXiv:2501.13956 , year=

  38. [56]

    The Thirteenth International Conference on Learning Representations , year=

    LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory , author=. The Thirteenth International Conference on Learning Representations , year=

  39. [57]

    arXiv preprint arXiv:2402.17753 , year=

    Evaluating very long-term conversational memory of llm agents , author=. arXiv preprint arXiv:2402.17753 , year=

  40. [58]

    International Conference on Learning Representations , volume=

    Forgetting transformer: Softmax attention with a forget gate , author=. International Conference on Learning Representations , volume=

  41. [59]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Memorybank: Enhancing large language models with long-term memory , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  42. [60]

    Proceedings of the 36th annual acm symposium on user interface software and technology , pages=

    Generative agents: Interactive simulacra of human behavior , author=. Proceedings of the 36th annual acm symposium on user interface software and technology , pages=

  43. [61]

    arXiv preprint arXiv:2601.02744 , year=

    SYNAPSE: Empowering LLM Agents with Episodic-Semantic Memory via Spreading Activation , author=. arXiv preprint arXiv:2601.02744 , year=

  44. [62]

    International Conference on Machine Learning , pages=

    A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts , author=. International Conference on Machine Learning , pages=. 2024 , organization=

  45. [63]

    2026 , eprint=

    FadeMem: Biologically-Inspired Forgetting for Efficient Agent Memory , author=. 2026 , eprint=

  46. [64]

    arXiv preprint arXiv:2604.00131 , year=

    Oblivion: Self-Adaptive Agentic Memory Control through Decay-Driven Activation , author=. arXiv preprint arXiv:2604.00131 , year=

  47. [65]

    , author=

    An integrated theory of the mind. , author=. Psychological review , volume=. 2004 , publisher=

  48. [66]

    , author=

    A temporal ratio model of memory. , author=. Psychological review , volume=. 2007 , publisher=

  49. [67]

    Advances in Neural Information Processing Systems , volume=

    H2o: Heavy-hitter oracle for efficient generative inference of large language models , author=. Advances in Neural Information Processing Systems , volume=

  50. [68]

    International Conference on Learning Representations , volume=

    Model tells you what to discard: Adaptive kv cache compression for llms , author=. International Conference on Learning Representations , volume=

  51. [69]

    Transactions on Machine Learning Research , year=

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author=. Transactions on Machine Learning Research , year=

  52. [70]

    Advances in Neural Information Processing Systems , volume=

    Learning to compress prompts with gist tokens , author=. Advances in Neural Information Processing Systems , volume=

  53. [71]

    arXiv preprint arXiv:2512.12818 , year=

    Hindsight is 20/20: Building agent memory that retains, recalls, and reflects , author=. arXiv preprint arXiv:2512.12818 , year=

  54. [72]

    Organization of memory , volume=

    Episodic and semantic memory , author=. Organization of memory , volume=. 1972 , publisher=

  55. [73]

    International conference on machine learning , pages=

    Neural episodic control , author=. International conference on machine learning , pages=. 2017 , organization=

  56. [74]

    International Conference on Learning Representations , year=

    Rapid Task-Solving in Novel Environments , author=. International Conference on Learning Representations , year=

  57. [75]

    Nature , volume=

    Hybrid computing using a neural network with dynamic external memory , author=. Nature , volume=. 2016 , publisher=

  58. [76]

    Behavioural Brain Research , volume=

    Episodic-like memory in animals , author=. Behavioural Brain Research , volume=. 2010 , publisher=

  59. [77]

    IEEE Transactions on Autonomous Mental Development , volume=

    Episodic-like memory for cognitive robots , author=. IEEE Transactions on Autonomous Mental Development , volume=. 2011 , publisher=

  60. [78]

    Towards virtual characters with a full episodic memory ii: The episodic memory strikes back , author=. Proc. empathic agents, AAMAS workshop , pages=

  61. [79]

    Proceedings of the National Academy of Sciences , volume=

    A taxonomy of prospection: Introducing an organizational framework for future-oriented cognition , author=. Proceedings of the National Academy of Sciences , volume=. 2014 , publisher=

  62. [80]

    Nature reviews neuroscience , volume=

    Remembering the past to imagine the future: the prospective brain , author=. Nature reviews neuroscience , volume=. 2007 , publisher=

  63. [81]

    Learning & Behavior , volume=

    Revisiting episodic-like memory in scrub jays: Is there more we can still learn from what--where--when caching behaviour? , author=. Learning & Behavior , volume=. 2025 , publisher=

  64. [82]

    Behavioural Brain Research , volume=

    Ten years of research into avian models of episodic-like memory and its implications for developmental and comparative cognition , author=. Behavioural Brain Research , volume=. 2010 , publisher=

  65. [83]

    ACM Computing Surveys (CSUR) , volume=

    Cache memories , author=. ACM Computing Surveys (CSUR) , volume=. 1982 , publisher=

  66. [84]

    International Conference on Machine Learning , pages=

    Fast inference from transformers via speculative decoding , author=. International Conference on Machine Learning , pages=. 2023 , organization=

  67. [85]

    arXiv preprint arXiv:2312.10997 , volume=

    Retrieval-augmented generation for large language models: A survey , author=. arXiv preprint arXiv:2312.10997 , volume=

  68. [86]

    arXiv preprint arXiv:2507.05257 , year=

    Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions , author=. arXiv preprint arXiv:2507.05257 , year=

  69. [87]

    Transactions on Machine Learning Research , year=

    Unsupervised Dense Information Retrieval with Contrastive Learning , author=. Transactions on Machine Learning Research , year=

  70. [88]

    arXiv preprint arXiv:2506.05176 , year=

    Qwen3 embedding: Advancing text embedding and reranking through foundation models , author=. arXiv preprint arXiv:2506.05176 , year=

  71. [89]

    Advances in Neural Information Processing Systems , volume=

    Llm evaluators recognize and favor their own generations , author=. Advances in Neural Information Processing Systems , volume=

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.