Pith. sign in

REVIEW 4 major objections 5 minor 46 references

MemoCue: Empowering LLM-Based Agents for Human Memory Recall via Strategy-Guided Querying

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that transforming a vague memory question into a cue-rich question—chosen by a recall strategy and refined by tree search—lets LLM agents help people recall memories much better than standard prompting.

desk verdict A genuinely new problem formulation and a thoughtful MCTS-plus-instruction-tuning pipeline for LLM-based memory recall assistance, but the headline claims rest on an evaluation loop that reuses the same simulated user that generated the training signal; the human study is too small to rescue it. read the letter →

arxiv 2507.23633 v1 pith:IMZR2T3A submitted 2025-07-31 cs.AI

Qian Zhao , Zhuo Sun , Bin Guo , Zhiwen Yu This is my paper
classification cs.AI
keywords MemoCuestrategy-guidedrecallLLMagentmemoryMonteCarloTreeSearch5WmapquerygenerationTip-of-the-Tongue
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

The paper tries to establish that an agent can help someone recall a forgotten memory by asking strategically chosen cue questions instead of trying to retrieve the memory from stored data. It introduces MemoCue, an LLM-based agent that maps a vague query such as "Where are my keys?" to one of five forgetting scenarios, picks one of fifteen recall strategies, and generates a cue-rich question designed to activate the memory. The authors report that MemoCue outperforms standard LLM prompting, zero-shot/few-shot, and chain-of-thought baselines on three long-term memory datasets, with a 17.74% gain in recall inspiration as rated by an LLM judge and higher win rates in small human evaluation. If the result holds, personal assistants could rely less on complete memory capture and more on guiding the user's own memory.

What carries the argument

The load-bearing mechanism is the strategy-guided cue query: the agent transforms an input query $Q_u$ into a cue query $Q_c$ via a strategy $s$, written $Q_c = T(Q_u, S)$, so that the user is prompted to activate memory associations rather than receive a direct answer. The Recall Router framework carries this out with two components: a 5W Recall Map that classifies queries into Event, Person, Location, Temporal, and Decision scenarios, and a hierarchical Monte Carlo Tree Search (SGR-MCTS) that selects among fifteen recall strategy patterns and generates cue queries. The reward for a cue is computed from simulated user feedback, combining recall accuracy (BERTScore against the true answer), recall focus (set-overlap distance on topic overlap), and recall depth (how many memory-element types appear in the response). This reward is what lets the system generate high-quality strategy responses without real human memory data.

What would settle it

A controlled human study with real participants given the same memory questions from these datasets, half receiving MemoCue's cue queries and half receiving baseline LLM responses, with recall success checked against known ground-truth answers; if MemoCue's recall-success rate is not meaningfully above baseline, or if the ordering of systems by real recall success does not match their BRS ordering, the paper's central claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that Tip-of-the-Tongue forgetting can be addressed by strategy-guided query transformation: rather than answering the user's incomplete query, the agent emits a cue query that activates associated memories. MemoCue does this by first classifying the query into one of five scenarios (Event, Person, Location, Temporal, Decision) with a 5W Recall Map, then selecting a strategy from a pool of fifteen patterns and generating a concrete cue question. The SGR-MCTS algorithm produces the training data by searching a hierarchical tree of strategies and cue queries, rewarding paths where a simulated user's response matches the true answer on BERTScore, topic overlap, and richness of memory elements. The resulting MemoStrategy instruction-tuning dataset is used to fine-tune several open-source LLMs, yielding MemoCue. On the PerLTQA, LoCoMo, and MemoryBank datasets, MemoCue achieves higher Balance of Recall Scores than the prompt-based baselines, and human evaluators prefer it over commercial closed-source models in 78–85% of comparisons.

Load-bearing premise

The entire training and evaluation loop assumes that an LLM acting as a simulated user remembers, forgets, and responds to cues the way a real person would; if real human recall behaves differently from this simulation, the optimized strategies and reported gains do not demonstrate real recall improvement.

Editorial extensions

If this is right

  • MemoCue's strategy prediction accuracy averages 90.36% across five open-source LLMs, so the 5W scenario classification and strategy pool transfer across model families without retraining the agent's core.
  • Because cue queries reduce reliance on stored memory, agents with incomplete memory banks can still offer useful recall support, changing the storage-versus-cuing tradeoff in personal assistants.
  • The SGR-MCTS reward design (accuracy, focus, depth) gives a concrete recipe for generating instruction-tuning data in other guidance tasks where direct human labels are scarce.
  • The BRS metric offers a way to score cue queries by balancing novelty from the original query against answer accuracy, applicable beyond this paper's datasets.

Reading between the lines

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

  • An implication the authors leave implicit: the same strategy taxonomy could support self-administered recall tools, where the user is handed the 5W map and strategies to generate their own cues, not just agent-generated ones.
  • The simulated-user reward could be replaced by an online loop of real user feedback, turning SGR-MCTS into a personalized cue optimizer that adapts strategies per individual; the paper does not test this.
  • Because BRS penalizes cue queries that resemble the original query, a cue that is novel but unhelpful could still score well if accuracy stays high; testing whether BRS predicts real human recall success would settle whether this is a weakness.
  • The 5W scenario taxonomy might transfer to interview or therapy settings, such as eyewitness questioning or reminiscence therapy, where guided cueing is known to matter; this is speculative.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes MemoCue, an LLM-based agent that assists human memory recall by transforming a user's original query into a cue-rich query via a structured strategy. The authors introduce a 5W Recall Map that classifies queries into five forgetting scenarios (Event, Person, Location, Temporal, Decision) and define fifteen recall strategy patterns. They then propose SGR-MCTS, a hierarchical Monte Carlo Tree Search algorithm that selects strategies and generates cue queries, rewarding queries that make a simulated LLM user (Qwen2.5-32B-Instruct) produce responses that match the ground-truth answer according to BERTScore. The resulting instructions are used to fine-tune several open-source LLMs, yielding MemoCue. Evaluation is conducted with a new metric BRS (Equation 10), an o1-preview LLM judge, and a small human study. The paper reports that MemoCue outperforms baselines by up to 11.24% in BRS, improves recall inspiration by 17.74% according to the LLM judge, and achieves 78-83% win rates in the human study. The central claim is that MemoCue significantly improves human memory recall.

Significance. If the central claim were validated, the paper would make a useful contribution to human-computer interaction and cognitive support: it formalizes a strategy-guided recall problem, proposes a plausible taxonomy of recall strategies grounded in memory theory, and offers a practical pipeline for generating cue queries. The authors also release the MemoStrategy dataset and fine-tune open-source models, which facilitates reproducibility and follow-up work. However, the significance is currently limited by the evaluation methodology. The automatic metric (BRS) and the MCTS training share the same simulated-user mechanism, and the human study does not measure actual recall success. The paper is therefore more convincing as a demonstration of a training and generation pipeline than as evidence that MemoCue improves real human memory recall.

major comments (4)
  1. [Section 3.3.2 / Equation (10) / Section 4.1.1] The automatic evaluation is circular with respect to the training objective. SGR-MCTS trains MemoCue by rewarding cue queries that cause the simulated user (Qwen2.5-32B-Instruct) to produce responses that match the true answer according to BERTScore. The BRS metric in Equation (10) scores the same kind of simulated responses with the same BERTScore-based accuracy term. The paper does not state explicitly that the evaluation uses the same simulator as training, but the methodology description implies this, and no calibration of the simulator against human memory behavior is provided. Consequently, the reported BRS gains and the 17.74% 'recall inspiration' improvement may simply reflect optimization in-distribution for the simulator, not improved recall for real users. The authors should either demonstrate that the simulated user's forgetting/remembering behavior matches human responses, or evaluate on a real human recall task.
  2. [Table 3 / Section 4.1 (Human Evaluation)] The human evaluation does not test memory recall success. Nine evaluators rank 30 pairs of responses on 'recall inspiration, response intelligence, and cue consistency,' which are subjective quality judgments about the generated cue queries. The evaluators are not asked to recall a to-be-remembered event, and there is no ground-truth memory target. Thus, the reported win rates (78-85%) do not support the claim that MemoCue 'improves human memory recall' or 'highlights its advantages in memory-recall applications.' A valid recall test would require participants to first encode a specific memory and then measure whether MemoCue's cues help them retrieve it.
  3. [Section 3.4 vs Appendix C.1] The reported size of the MemoStrategy dataset is internally inconsistent. Section 3.4 states that the dataset contains 5,805 samples (5,200 training / 605 testing), while Appendix C.1 reports 4,805 samples (4,500 training / 305 testing). This contradiction must be resolved because the dataset is a central contribution and its exact composition is required for reproducibility.
  4. [Section 3.3.2 / Equation (8)] The paper defines three reward components (Recall Accuracy, Recall Focus, Recall Depth) but does not specify how they are combined into the final reward Rend used in backpropagation. For example, are they summed, weighted, or multiplied? Without this information, the SGR-MCTS algorithm is not fully reproducible, and the relative contribution of each component cannot be assessed.
minor comments (5)
  1. [Throughout] There are several typos and minor errors: 'devided' in Section 3.1, 'exprimental' in Section 4.1, 'LlaMA' instead of 'Llama' in Section 4.1, and 'ReBERTa' instead of 'RoBERTa' in Appendix B.
  2. [Equation (10)] BRS uses BERTScore as an accuracy measure and cosine similarity as a penalty, but the paper does not state which BERTScore variant is used, how the similarity is computed (e.g., which embedding model), or how α=0.3 was selected. These details are necessary for reproducibility.
  3. [Figure 3 caption] The caption states that examples are 'shown in the right,' but the examples are presented in a table, not on the right side of the figure. Please correct the caption.
  4. [Section 4.1 / Datasets] For PerLTQA, the paper states that the dataset is transformed into 35,179 memory streams by splitting long query entries, but it does not specify how this transformation is performed or whether the split query entries preserve the original answer mapping. This is important for reproducibility.
  5. [Appendix B] The classification model is trained on 2,000 samples annotated by GPT-3.5-turbo, but the paper reports only aggregate accuracy in Figure 4. It would be helpful to report per-scenario precision/recall and the number of annotation disagreements resolved during manual verification.

Circularity Check

2 steps flagged · score 6.0 of 10

The automatic BRS evaluation reuses the same simulated-user/BERTScore objective that generated MemoCue's training data, so the reported automatic gains are in-distribution; the 17.74% headline figure comes from an o1-preview judge, not from any human recall test.

  1. fitted input called prediction [Section 3.3.2 (Eq. 5) and Section 4.1.1 (Eq. 10)]
    "After the simulated user generates response rt based on Qc ... send = ( Success, if acc (rt, rans) ≥ θacc ... where acc(·) is the BERTScore based similarity metric ... BRS = Acc(rres,i, rans,i) / (1 + α · Sim(Qu,i, Qc,i)) where Acc (·, ·) is the accuracy of response calculated by BERTScore"

    The same Qwen2.5-32B-Instruct simulated user produces the response in both SGR-MCTS training and BRS evaluation, and both stages score that response against the true answer with BERTScore. SGR-MCTS therefore selects and backpropagates rewards for cue queries that maximize the simulated user's BERTScore similarity, and BRS then reports that same kind of quantity as an independent 'balance of recall score'. MemoCue's BRS improvement is by construction aligned with the optimization target, so the automatic evaluation does not provide independent evidence about human memory recall.

  2. fitted input called prediction [Appendix C.1, Data Quality Optimization]
    "we delete invalid samples that failed to successfully activate user memory-recall, that is, scenarios where user responses were unclear or failed to provide valid information... marked as failed samples and remove from the dataset."

    The MemoStrategy training corpus is built by retaining only simulated-user trajectories that satisfied the BERTScore-based success criterion used in SGR-MCTS. MemoCue is then fine-tuned on this filtered corpus and evaluated with the same simulator and BERTScore-style accuracy. The dataset is thus selected to make the model perform well on the evaluation metric, and no external human recall signal enters the loop.

full rationale

The paper's main automatic evaluation chain is partially circular. Both the SGR-MCTS reward mechanism (Section 3.3.2, Eq. 5 and Eq. 7) and the BRS metric (Section 4.1.1, Eq. 10) rely on a Qwen2.5-32B-Instruct simulated user producing a response to the cue query, scored by BERTScore against the true answer. MemoCue is fine-tuned on cue queries selected to maximize that exact simulated-user BERTScore, so the reported BRS improvements in Tables 1-2 are in-distribution for the optimization target rather than a fresh measurement. The paper honestly states 'due to the lack of real-world human memory recall datasets, it is difficult to directly evaluate the end-to-end recall success performance in this work', which supports the interpretation that BRS is a proxy. The headline 17.74% 'recall inspiration' gain comes from an independent o1-preview judge, so the central claim is not entirely reducible to the training objective; however, that judge is an LLM, not a human, and it rates subjective qualities rather than measuring whether recall actually succeeded. The small human study (9 evaluators, 30 pairs, win rates 83-85%) uses subjective ranking rather than recall success. A secondary but notable issue is the internal dataset-size inconsistency: Section 3.4 reports 5805 samples (5200/605) while Appendix C.1 reports 4805 samples (4500/305), undermining reproducibility. Overall, the automatic evaluation loop is circular at its core, but the presence of an independent LLM judge and a small human preference study keeps the paper from being fully self-justifying. Score 6 reflects partial circularity of the main evaluation chain.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim relies on the assumptions that LLM-simulated users behave like real humans in recall, that BERTScore similarity is a good proxy for successful recall, and that the hand-designed 5W strategy taxonomy covers real forgetting scenarios. The free parameters (alpha, MCTS constants, thresholds) are chosen by the authors and affect the reported scores. No physical entities are invented; the 5W Recall Map and 15 strategy patterns are conceptual constructs that are part of the method design rather than new ontological claims.

free parameters (8)
  • alpha in BRS metric = 0.3
    Penalty weight in the BRS metric (Eq. 10) chosen by hand; all reported BRS scores depend on this value.
  • MCTS exploration constant c = not specified
    UCT exploration constant (Eq. 4) not reported; impacts strategy search and final results.
  • epsilon-greedy schedule = start at 1, decrease by 0.05
    Exploration factor in simulation phase, chosen without sensitivity analysis.
  • discount factor gamma = not specified
    Discount factor for backpropagation (Eq. 8); affects reward propagation.
  • MCTS iterations T = 120
    Number of MCTS iterations; Figure 6 shows performance peaks at 120 but drops at 180, indicating sensitivity.
  • top-k paths k = 5
    Number of best paths collected for dataset construction.
  • accuracy threshold theta_acc = not specified
    Success threshold for memory recall in terminal condition (Eq. 5).
  • max turns N_max = not specified
    Maximum dialog turns before failure in terminal condition (Eq. 5).
assumptions (5)
  • domain assumption Human forgetting is often caused by ineffective memory activation rather than loss of memory.
    Cited from memory theories (Collins & Loftus 1975; Meyer & Bock 1992) in the Introduction; the entire SGR problem formulation depends on this.
  • domain assumption An LLM can simulate a user's memory recall process accurately enough to guide strategy selection.
    SGR-MCTS uses Qwen2.5-32B-Instruct as simulated user (Section 3.3.2); if the simulation diverges from real recall, the optimized strategies and the BRS evaluation are invalid.
  • ad hoc to paper BERTScore similarity between a generated response and the true answer is a valid measure of recall success.
    Used in the terminal condition (Eq. 5) and the BRS metric (Eq. 10); no evidence that BERTScore reflects actual human recall.
  • ad hoc to paper The 5W taxonomy and fifteen strategies cover the space of forgetting scenarios.
    Defined by the authors in Table 6; no external validation that these categories are exhaustive or optimal.
  • domain assumption o1-preview with few-shot prompts provides a reliable evaluation of recall inspiration and other qualities.
    Used as the LLM evaluator in Section 4.1.1; LLM judges can be biased despite the BPC strategy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MemoCue: Empowering LLM-Based Agents for Human Memory Recall via Strategy-Guided Querying." pith.science (2026). https://pith.science/paper/IMZR2T3A

@misc{pith2026250723633,
  author       = {Pith},
  title        = {Pith review of: MemoCue: Empowering LLM-Based Agents for Human Memory Recall via Strategy-Guided Querying},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IMZR2T3A}},
  note         = {Machine review of arXiv:2507.23633}
}
read the original abstract

Agent-assisted memory recall is one critical research problem in the field of human-computer interaction. In conventional methods, the agent can retrieve information from its equipped memory module to help the person recall incomplete or vague memories. The limited size of memory module hinders the acquisition of complete memories and impacts the memory recall performance in practice. Memory theories suggest that the person's relevant memory can be proactively activated through some effective cues. Inspired by this, we propose a novel strategy-guided agent-assisted memory recall method, allowing the agent to transform an original query into a cue-rich one via the judiciously designed strategy to help the person recall memories. To this end, there are two key challenges. (1) How to choose the appropriate recall strategy for diverse forgetting scenarios with distinct memory-recall characteristics? (2) How to obtain the high-quality responses leveraging recall strategies, given only abstract and sparsely annotated strategy patterns? To address the challenges, we propose a Recall Router framework. Specifically, we design a 5W Recall Map to classify memory queries into five typical scenarios and define fifteen recall strategy patterns across the corresponding scenarios. We then propose a hierarchical recall tree combined with the Monte Carlo Tree Search algorithm to optimize the selection of strategy and the generation of strategy responses. We construct an instruction tuning dataset and fine-tune multiple open-source large language models (LLMs) to develop MemoCue, an agent that excels in providing memory-inspired responses. Experiments on three representative datasets show that MemoCue surpasses LLM-based methods by 17.74% in recall inspiration. Further human evaluation highlights its advantages in memory-recall applications.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 28 canonical work pages

  1. [2]

    Baddeley

    A. Baddeley. Working memory. Science, 255(5044):556–559, 1992

    work page 1992

  2. [3]

    tip of the tongue

    R. Brown and D. McNeill. The “tip of the tongue” phenomenon. Jour- nal of verbal learning and verbal behavior, 5(4):325–337, 1966

    work page 1966

  3. [4]

    D. M. Burke, D. G. MacKay, J. S. Worthley, and E. Wade. On the tip of the tongue: What causes word finding failures in young and older adults? Journal of memory and language, 30(5):542–579, 1991

    work page 1991

  4. [5]

    Chandra, S

    S. Chandra, S. Sharma, R. Chaudhuri, and I. Fiete. Episodic and as- sociative memory from spatial scaffolds in the hippocampus. Nature, pages 1–13, 2025

    work page 2025

  5. [6]

    A. M. Collins and E. F. Loftus. A spreading-activation theory of seman- tic processing. Psychological review, 82(6):407, 1975

    work page 1975

  6. [7]

    F. I. Craik and R. S. Lockhart. Levels of processing: A framework for memory research. Journal of verbal learning and verbal behavior , 11 (6):671–684, 1972

    work page 1972

  7. [8]

    Q. Dong, L. Li, D. Dai, C. Zheng, J. Ma, R. Li, H. Xia, J. Xu, Z. Wu, T. Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022

    arXiv 2022

  8. [9]

    Y . Du, H. Wang, Z. Zhao, B. Liang, B. Wang, W. Zhong, Z. Wang, and K.-F. Wong. Perltqa: A personal long-term memory dataset for memory classification, retrieval, and synthesis in question answering. arXiv preprint arXiv:2402.16288, 2024

    arXiv 2024

Show all 46 references
  1. [10]

    Eichenbaum, P

    H. Eichenbaum, P. Dudchenko, E. Wood, M. Shapiro, and H. Tanila. The hippocampus, memory, and place cells: is it spatial memory or a memory space? Neuron, 23(2):209–226, 1999

    1999

  2. [11]

    D. D. Georgiev, I. Georgieva, Z. Gong, V . Nanjappan, and G. V . Georgiev. Virtual reality for neurorehabilitation and cognitive enhance- ment. Brain sciences, 11(2):221, 2021

    2021

  3. [12]

    B. J. Gutiérrez, Y . Shu, Y . Gu, M. Yasunaga, and Y . Su. Hipporag: Neu- robiologically inspired long-term memory for large language models. arXiv preprint arXiv:2405.14831, 2024

    2024 arXiv

  4. [13]

    R. S. Herz. The effects of cue distinctiveness on odor-based context- dependent memory. Memory & cognition, 25:375–380, 1997

    1997

  5. [14]

    E. J. Hu, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen, et al. Lora: Low-rank adaptation of large language models. In Interna- tional Conference on Learning Representations, 2021

    2021

  6. [15]

    Hurst, A

    A. Hurst, A. Lerer, A. P. Goucher, A. Perelman, A. Ramesh, A. Clark, A. Ostrow, A. Welihinda, A. Hayes, A. Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  7. [16]

    P. Jaccard. The distribution of the flora in the alpine zone. 1. New phytologist, 11(2):37–50, 1912

    1912

  8. [17]

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. d. l. Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023

    2023 arXiv

  9. [18]

    Kaiya, M

    Z. Kaiya, M. Naim, J. Kondic, M. Cortes, J. Ge, S. Luo, G. R. Yang, and A. Ahn. Lyfe agents: Generative agents for low-cost real-time social interactions. arXiv preprint arXiv:2310.02172, 2023

    2023 arXiv

  10. [19]

    Kandappu, V

    T. Kandappu, V . Subbaraju, and Q. Xu. Privacyprimer: Towards privacy-preserving episodic memory support for older adults. Proceed- ings of the ACM on Human-Computer Interaction , 5(CSCW2):1–32, 2021

    2021

  11. [20]

    A. T. Khan, S. Li, and X. Cao. Human guided cooperative robotic agents in smart home using beetle antennae search.Science China Information Sciences, 65(2):122204, 2022

    2022

  12. [21]

    Kocsis and C

    L. Kocsis and C. Szepesvári. Bandit based monte-carlo planning. In European conference on machine learning , pages 282–293. Springer, 2006

    2006

  13. [22]

    Kojima, S

    T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa. Large lan- guage models are zero-shot reasoners. Advances in neural information processing systems, 35:22199–22213, 2022

    2022

  14. [23]

    H. D. Lasswell. The structure and function of communication in society. The communication of ideas, 37(1):136–139, 1948

    1948

  15. [24]

    Y . Li, Y . Zhang, and L. Sun. Metaagents: Simulating interactions of human behaviors for llm-based task-oriented coordination via collabo- rative generative agents. arXiv preprint arXiv:2310.06500, 2023

    2023 arXiv

  16. [25]

    Y . Liu. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 364, 2019

    1907 arXiv

  17. [26]

    J. Lu, S. An, M. Lin, G. Pergola, Y . He, D. Yin, X. Sun, and Y . Wu. Memochat: Tuning llms to use memos for consistent long-range open- domain conversation. 2023

    2023

  18. [27]

    Maharana, D.-H

    A. Maharana, D.-H. Lee, S. Tulyakov, M. Bansal, F. Barbieri, and Y . Fang. Evaluating very long-term conversational memory of llm agents. arXiv preprint arXiv:2402.17753, 2024

    2024 arXiv

  19. [28]

    C. B. Martin, B. Hong, R. N. Newsome, K. Savel, M. E. Meade, A. Xia, C. J. Honey, and M. D. Barense. A smartphone intervention that en- hances real-world memory and promotes differentiation of hippocam- pal activity in older adults. Proceedings of the National Academy of Scien...

    2022

  20. [29]

    A. Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024

    2024

  21. [30]

    A. S. Meyer and K. Bock. The tip-of-the-tongue phenomenon: Blocking or partial activation? Memory & Cognition, 20(6):715–726, 1992

    1992

  22. [31]

    Naveed, A

    H. Naveed, A. U. Khan, S. Qiu, M. Saqib, S. Anwar, M. Usman, N. Akhtar, N. Barnes, and A. Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435, 2023

    2023 arXiv

  23. [32]

    R. OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2(5), 2023

    2023 arXiv

  24. [33]

    Sarfraz, E

    F. Sarfraz, E. Arani, and B. Zonooz. Sparse coding in a dual memory system for lifelong learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 9714–9722, 2023

    2023

  25. [34]

    Sarthi, S

    P. Sarthi, S. Abdullah, A. Tuli, S. Khanna, A. Goldie, and C. D. Man- ning. Raptor: Recursive abstractive processing for tree-organized re- trieval. In The Twelfth International Conference on Learning Represen- tations, 2024

    2024

  26. [35]

    Schindler, R

    S. Schindler, R. V ormbrock, and J. Kissler. Encoding in a social feed- back context enhances and biases behavioral and electrophysiological correlates of long-term recognition memory. Scientific Reports, 12(1): 3312, 2022

    2022

  27. [36]

    E. Song, W. Chai, G. Wang, Y . Zhang, H. Zhou, F. Wu, H. Chi, X. Guo, T. Ye, Y . Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pages 18221–18232, 2024

    2024

  28. [37]

    L. R. Squire. Mechanisms of memory. Science, 232(4758):1612–1619, 1986

    1986

  29. [38]

    P. Wang, L. Li, L. Chen, Z. Cai, D. Zhu, B. Lin, Y . Cao, Q. Liu, T. Liu, and Z. Sui. Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926, 2023

    2023 arXiv

  30. [39]

    W. Wang, L. Dong, H. Cheng, X. Liu, X. Yan, J. Gao, and F. Wei. Aug- menting language models with long-term memory. Advances in Neural Information Processing Systems, 36:74530–74543, 2023

    2023

  31. [40]

    J. Wei, M. Bosma, V . Y . Zhao, K. Guu, A. W. Yu, B. Lester, N. Du, A. M. Dai, and Q. V . Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021

    2021 arXiv

  32. [41]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  33. [42]

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  34. [43]

    Young, B

    A. Young, B. Chen, C. Li, C. Huang, G. Zhang, G. Zhang, H. Li, J. Zhu, J. Chen, J. Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024

    2024 arXiv

  35. [44]

    Zhang, V

    T. Zhang, V . Kishore, F. Wu, K. Q. Weinberger, and Y . Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations, 2019

    2019

  36. [45]

    Zhang, X

    Z. Zhang, X. Bo, C. Ma, R. Li, X. Chen, Q. Dai, J. Zhu, Z. Dong, and J.- R. Wen. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501, 2024

    2024 arXiv

  37. [46]

    Zhong, Z

    T. Zhong, Z. Liu, Y . Pan, Y . Zhang, Y . Zhou, S. Liang, Z. Wu, Y . Lyu, P. Shu, X. Yu, et al. Evaluation of openai o1: Opportunities and chal- lenges of agi. arXiv preprint arXiv:2409.18486, 2024

    2024

  38. [47]

    When did I take my medication?

    W. Zhong, L. Guo, Q. Gao, H. Ye, and Y . Wang. Memorybank: En- hancing large language models with long-term memory. In Proceed- ings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19724–19731, 2024. A Definitions of Memory-Recall Strategies We define 15 me...

    2024

Pith tools

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