CAMeR: Keyword-Gated Hybrid Activation for Adaptive Memory Retention in LLM Agents

Haowen Lai

REVIEW 5 major objections 4 minor 16 references

CAMeR shows that adding a keyword-overlap term to embedding similarity makes LLM-agent memory retention sharply discriminate relevant from irrelevant memories, with a 1.6× larger retention gap and 83.2% token savings.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 13:55 UTC pith:Z3W2QPHM

load-bearing objection Useful hybrid-gating idea and a needed benchmark, but the printed numbers don't add up—code and fixes required before the 1.6x claim can be trusted. the 5 major comments →

arxiv 2607.20458 v1 pith:Z3W2QPHM submitted 2026-05-15 cs.CL cs.AI

CAMeR: Keyword-Gated Hybrid Activation for Adaptive Memory Retention in LLM Agents

Haowen Lai This is my paper

classification cs.CL cs.AI

keywords keyword-gated hybrid activationadaptive memory retentionLLM agentsJaccard similarityembedding cosinescissors gapCAMeR-Benchtoken savings

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that LLM agents can remember adaptively by pairing an embedding-similarity gate with a word-level keyword-overlap term, and by reinforcing memories that pass the combined threshold while decaying the rest. On a 76-memory, 100-round synthetic benchmark the keyword gate produces a retention gap 1.6 times larger than embedding-only gating, keeps frequently used memories at substantially higher weight than never-referenced ones, and saves 83.2% of tokens versus full-context injection. The authors isolate the keyword gate—not the learnable decay component—as the main driver of differentiation, and show that purely time-driven forgetting curves collapse over long horizons. A sympathetic reader would care because the mechanism is simple, zero-shot, and directly addresses the false-positive activation problem in embedding-only memory systems.

Core claim

The central discovery is that a hybrid activation score—0.6 times embedding cosine plus 0.4 times keyword Jaccard over extracted keywords—produces weight signals that separate high-frequency memories from never-referenced ones far better than embedding similarity alone. On the paper's CAMeR-Bench benchmark the scissors gap (mean weight of activated minus never-activated memories) reaches 0.039 for the full hybrid gate versus 0.024 for embedding-only, a 1.6× improvement; per-cluster weights track query frequency with Spearman ρ = 0.90. The keyword term works by exploiting sparsity: roughly 90% of memory-query pairs share no keywords, so the Jaccard term stretches the score distribution and su

What carries the argument

The key mechanism is the keyword-gated hybrid activation function: score(mi,q) = α·cos(e(mi),e(q)) + (1−α)·Jaccard(K(mi),K(q)) with α=0.6 and threshold τ=0.25, followed by an asymmetric reinforcement-decay update: weight ← clamp(weight·0.99 + 0.2) if activated, else weight ← weight·0.99. The Jaccard term over keyword sets acts as a precision filter that rejects semantically adjacent but factually irrelevant memories, while multiplicative decay plus additive reinforcement creates bounded, order-dependent weight dynamics. A long-term migration mechanism switches frequently accessed memories to a slower decay rate after about 50 activations.

Load-bearing premise

The entire result rests on CAMeR-Bench's synthetic structure—template memories and queries grouped into 8 topic clusters with a graded query-frequency gradient and a zero-query background—being representative of real LLM-agent deployments, and on the scissors gap capturing genuine retention quality rather than just the gate's own binary choices.

What would settle it

Run CAMeR on a naturally occurring long-dialogue corpus with human-labeled relevance judgments; if high-frequency, genuinely relevant memories do not end up with higher weights than low-frequency irrelevant ones (i.e., the scissors gap computed against ground-truth labels is near zero or negative), the core claim fails. Alternatively, compute the Jaccard-overlap distribution on a large real conversation set; if far more than ~10% of memory-query pairs share keywords, the sparsity mechanism that drives the gate's precision advantage collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

If CAMeR's hybrid gate is correct, embedding-only memory systems can be improved by a zero-shot keyword term with no learned parameters, reducing false-positive activations.
The 83.2% token savings (39k vs 231k cumulative) means a top-5 weight-augmented retrieval can substitute for full-context injection in long dialogues at a fraction of the cost.
Fixed-rate decay with conditional reinforcement outperforms both learned per-memory decay (at 81 parameters, which underfits) and exponential/Ebbinghaus time-decay curves, which collapse to near-zero over 100 rounds.
The weight signals from the keyword gate carry sufficient information to improve retrieval precision (+0.008), a causal chain absent in embedding-only and no-gate conditions.
CAMeR-Bench provides a controlled testbed with a graded activation gradient and zero-query background, filling a gap left by retrieval-focused benchmarks like LoCoMO and LongMemEval.

Where Pith is reading between the lines

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

The sparsity argument suggests the Jaccard term's benefit grows with vocabulary diversity; in real-world long-tail corpora the gate may separate memories even more sharply than in this synthetic benchmark, but the paper leaves this untested.
One testable extension is replacing the fixed α=0.6 with a per-query or per-cluster learned weighting, which could adapt the embedding/keyword balance to dialogue tempo or domain specificity; the paper notes static hyperparameters as a limitation.
The 'scissors gap' as defined measures differentiation between memories the gate activated vs. never activated, so it partly reflects the gate's own selectivity; a fairer ecological test would measure retention against ground-truth relevance labels in real dialogues, which the paper calls for in §6.4.4.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Useful hybrid-gating idea and a needed benchmark, but the printed numbers don't add up—code and fixes required before the 1.6x claim can be trusted.

read the letter

Dear Colleague,

The punchline: the keyword+embedding hybrid gate is a reasonable idea, and CAMeR-Bench addresses a real gap in how we evaluate retention. But as printed, the evidence doesn't support the headline claims—there are arithmetic inconsistencies and a missing code release that need fixing before I'd trust the numbers.

What's genuinely new: combining a word-level Jaccard gate with embedding cosine for activation, then using the resulting weights to modulate retrieval, is a clean idea that prior memory systems (Mem0, Memory-R1, Oblivion) don't do. The benchmark design—8 topic clusters with graded query frequencies and a zero-query background—is a real contribution, because LoCoMO and LongMemEval can't produce an activation gradient. The ablation story is coherent: the gate, not the learned MLP decay, is doing the work at this scale.

The soft spots are real. First, the motivating example in Table 2 contradicts itself: the two memories marked 'suppressed' have hybrid scores 0.31 and 0.35, both above the stated threshold τ=0.25. That example is the paper's core intuition and it fails on its own arithmetic. Second, the background weights in Table 4 (0.605) don't match the decay schedule. A never-accessed memory injected during rounds 1–20 should sit around 0.38–0.45 after 100 rounds of 0.99^t decay. The 0.605 value implies cross-cluster activation, which means those memories aren't actually 'never-activated' and the scissors gap is measuring something noisier than the paper claims. Figure 2 giving Cluster H at 0.483 only deepens the inconsistency. Third, the 76-to-67 memory pool drop is unexplained, and there's no code or data link despite the release statement.

None of this kills the idea—it's fixable. But with all results from a single seed and small effect sizes (0.039 vs 0.024, +0.008 precision), the current evidence is not enough. Re-run multiple seeds, fix the arithmetic, reconcile the weights, and release the code and benchmark.

Who should read it: people building LLM-agent memory systems, and anyone thinking about symbolic/neural hybrid signals. A serious referee should see it, but only after the authors do a major revision and provide the artifacts. I wouldn't desk-reject it; I'd send it back for evidence.

Referee Report

5 major / 4 minor

Summary. The paper proposes CAMeR, a memory retention framework for LLM agents that scores memories by a weighted combination of embedding cosine and keyword Jaccard similarity, reinforces memories that exceed a threshold, and applies multiplicative decay to all others, with an optional 81-parameter MLP per-memory decay schedule. The authors introduce CAMeR-Bench, a synthetic 76-memory, 100-round benchmark with eight topic clusters and a graded query-frequency gradient, and report that the keyword gate produces a scissors gap of 0.039 versus 0.024 for embedding-only gating (a 1.6× improvement), that top-5 retrieval saves 83.2% of tokens relative to full-context injection while improving retrieval precision by +0.008, and that ablations identify the keyword gate—not learnable decay—as the primary performance driver.

Significance. If correct, the paper offers a simple, interpretable, and low-cost mechanism for improving relevance-based memory retention, and CAMeR-Bench would fill a real gap in controlled evaluation of retention dynamics. The ablation design is thoughtful, particularly the separation of gating, decay, and retrieval effects, and the paper includes several useful external checks (per-cluster gradient, retrieval precision gain, token consumption). However, the central numerical claim currently rests on a single-seed run whose principal table is internally inconsistent with the stated decay dynamics, and the motivating example contains a threshold arithmetic error. The overall idea is promising, but the evidence as printed is not yet sufficient to support the headline conclusions.

major comments (5)

[§5.1, Table 4 and §3.2, Eq. (2); Table 3] The reported w_bg = 0.605 for CAMeR-Full is inconsistent with the published dynamics. Under Eq. (2) with γ = 0.99, a never-activated memory injected in Cluster H (injection window rounds 1–20, Table 3) must satisfy w_bg ≤ 0.99^80 ≈ 0.45; the expected injection round gives ≈0.38. The value 0.605 corresponds to roughly 50 rounds of decay, implying injection near round 50, outside the stated window. The §5.2 'semantic adjacency' explanation would require a_i > 0 for those memories, removing them from the never-activated class by the paper's own definition (Algorithm 1, line 8). Thus the headline 0.039 vs. 0.024 comparison is unsupported as printed.
[§5.2 (Fig. 2) vs. §5.1 (Table 4)] There is a direct numerical contradiction between Figure 2, which reports Cluster H's average final weight as 0.483, and Table 4, which reports w_bg = 0.605 for the same condition and the same 100-round run. Both exhibits purportedly describe the never-accessed Cluster H memories. The two numbers cannot both be correct. Please identify the source of the discrepancy and re-run the experiments before the scissors-gap comparisons can be evaluated.
[§4.3, §3.1, Algorithm 1] The scissors-gap metric is partly self-referential: it compares weights of memories the gate itself labels activated (a_i > 0) with weights of memories the gate labels background (a_i = 0). Because the gate's own threshold determines these classes, a stricter gate can mechanically increase the gap without improving true retention. The per-cluster gradient (Fig. 2) is a good external anchor, but it is currently entangled with the same contamination issue and the single-seed run. Please report the per-cluster gradient for all conditions, provide confidence intervals or multiple seeds, and, if possible, tie the scissors gap to an independent relevance label.
[§4.4, Tables 4–5] All experiments use seed 42 with no variance estimates or significance tests. The key comparisons—0.039 vs. 0.024 scissors gap and +0.008 vs. −0.002 precision gain—are point estimates from a single run and could plausibly be noise. Given the small number of memories (76) and the modest effect sizes, multiple seeds (or bootstrap resampling) and a significance test are needed before the 1.6× and precision-gain claims can be accepted.
[§3.4, Table 2] The qualitative case study is internally inconsistent. For the two memories marked ✗ as suppressed, the hybrid scores are 0.6×0.47 + 0.4×0.07 = 0.31 and 0.6×0.52 + 0.4×0.09 = 0.35, both strictly greater than τ = 0.25. The text's claim that 'with the hybrid gate, only one exceeds τ = 0.25' is therefore false. The example as printed does not demonstrate false-positive suppression; it demonstrates the opposite. Please correct the scores, the threshold, or the narrative.

minor comments (4)

[§4.2 and §5.4] The 'Phase 1' dataset is used before it is defined. Provide an explicit description of Phase 1 and its relationship to the 100-round CAMeR-Bench results in §4.
[Tables 4 and 6] Clarify which configuration is the primary model. Table 4 uses CAMeR-Full (with MLP) for the headline result, while Table 6 says '−MLP (our primary)' is the primary variant. State explicitly which configuration the abstract and conclusion refer to.
[§6.1] The claims that random sentence pairs share a keyword with probability 0.08–0.12 and that cosine similarity rarely falls below 0.20–0.25 are not supported by a reference, table, or appendix. Consider adding a short empirical characterization of the score distributions.
[References] Several references (e.g., [2], [4]) are arXiv preprints or non-archival heuristics with 2026 dates. Verify bibliographic details and peer-reviewed status as appropriate for the journal.

Circularity Check

1 steps flagged

Scissors-gap metric is self-defined by CAMeR's own gate, partially circular; independent anchors keep score at 4.

specific steps

self definitional [§3.2 Eq. (2), §4.3 'Scissors Gap', §5.1 Finding 1 (Table 4)]
"Scissors Gap(↑): Mean weight of activated (ai > 0) memories minus mean weight of never-activated (ai = 0) background memories. This is our primary metric — it directly measures the system’s ability to differentiate relevant from irrelevant information. … CAMeR-Full achieves scissors gap 0.039, 1.6× larger than EmbOnly’s 0.024."

The two classes in the metric are generated by the same gate the metric is meant to validate: 'activated' means ai > 0, i.e. the gate's own thresholding (Eq. 1, τ = 0.25). Eq. (2) then mechanically adds +0.2 to exactly those memories and applies 0.99 decay to the rest. After 100 rounds the gap between 'activated' and 'never-activated' weights is therefore the gate's selectivity written into the weights; a stricter gate trivially widens the gap. The abstract's framing as 'between high-frequency and never-referenced memories' substitutes ground-truth labels for the gate's self-assigned labels, which is valid only if the gate is already assumed correct. The per-cluster gradient against ground-truth frequencies (ρ=0.90) and the weight-augmented precision gain provide independent anchoring, so

full rationale

The main circularity is the scissors-gap metric: the classes 'activated' and 'never-activated' are produced by CAMeR's own gate, and the update rule then reinforces exactly the activated memories and decays the others. Consequently the headline 1.6× gap is largely a self-consistency artifact of the gate's selectivity, not an independent measure of retention quality. This is only partial because the paper also supplies externally anchored evidence: the per-cluster weight gradient is compared with the benchmark's true query frequencies (Spearman ρ = 0.90, p < 0.01) and the weight-augmented precision gain (+0.008) is evaluated against pure cosine ranking, both of which do not depend on the gate's own labels. No load-bearing self-citation appears; all references are to external work. A separate reproducibility concern, not counted as circularity, is that Table 4's background weight w_bg = 0.605 for Cluster H is hard to reconcile with the stated 0.99 decay over the rounds 1–20 injection window (expected ≈0.38–0.45 by round 100); this appears to be a correctness/reproducibility issue. The score of 4 reflects partial circularity of the primary metric with meaningful independent anchors preventing a higher score.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 1 invented entities

CAMeR contributes a hand-specified dynamical system: six fixed hyperparameters (α, τ, γ, Δw, γ_long, top-k=5) plus an optional 81-parameter MLP trained on heuristic pseudo-labels. The weight dynamics are fully explicit (Eq. 2), but the benchmark's representativeness and the metric's self-referential activation labels are assumptions pulled from the paper's design rather than demonstrated externally. No new physical entities are postulated; the long-term accumulator is a mechanism variable with no out-of-bench evidence.

free parameters (8)

α (hybrid gate weight) = 0.6
Relative weight of embedding cosine vs. keyword Jaccard in Eq. 1 (§3.1); hand-set, stated as fixed.
τ (activation threshold) = 0.25
Threshold on the hybrid score that decides which memories are reinforced and which are counted as activated (§3.1); hand-set.
γ (decay rate) = 0.99
Multiplicative decay per non-activated round (§3.2); hand-set.
Δw (reinforcement increment) = 0.2
Additive boost on activation (§3.2); hand-set, roughly 20× single-round decay.
γ_long (long-term decay) = 0.995
Slower decay after long-term migration (§3.3); hand-set; never triggered within 100 rounds (~50 activations required).
MLP architecture and pseudo-label training setup = 3-16-1, 81 params; 200 MSE + 100 ranking epochs on ~280 samples
Learned per-memory decay (§3.6, §4.4) trained on Phase 1 trajectories and evaluated on Phase 1; underfits, per §5.4.
Baseline parameters (Oblivion λ, SuperLocal S0) = λ=0.01; S0=5
Chosen so that after 100 rounds the formulas produce near-zero weights by construction (e^{−1}≈0.37; e^{−20}) (§4.2, §5.1).
Phase-1 ablation design = 14 memories, 20 rounds, 4 clusters
The 8-condition ablation runs on this smaller dataset (§5.4), which shapes the conclusion that the gate, not the MLP, drives performance.

axioms (6)

domain assumption KeyBERT (with YAKE/TF-IDF fallback) extracts keyword sets whose overlap faithfully tracks topical relevance
The gate's precision advantage rests entirely on keyword sets being reliable relevance signals (§3.1, §3.4, §6.1); no intrinsic evaluation of extraction quality is given, and the authors flag language dependence (§6.4.2).
domain assumption Sparsity prior: random English memory-query pairs share keywords with probability ~0.08–0.12, so Jaccard ≈ 0 for ~90% of pairs
Empirical claim from 'our corpus' in §6.1; underpins the explanation of why the gate stretches the score distribution, but no distribution or sample size is provided.
domain assumption Embedding cosine dynamic range is compressed (rarely below ~0.20–0.25 for arbitrary English pairs)
Underpins the thresholding-failure narrative in §6.1; stated without data or citation.
domain assumption CAMeR-Bench's synthetic structure (template memories/queries, graded query frequencies, zero-query background) is a valid stand-in for real adaptive-retention workloads
All headline results come from this benchmark (§4.1); the authors acknowledge in §6.4.4 that deployment in real dialogues is needed for ecological validation.
ad hoc to paper The scissors-gap classes (activated a_i>0 vs. never-activated a_i=0), assigned by the gate under test, validly measure retention quality
The metric's positive/negative split is decided by the same gate being evaluated (§4.3, §5.1), making the primary metric partly self-referential; corroborated only indirectly by the per-cluster gradient and precision gain.
domain assumption all-MiniLM-L6-v2 embedding cosine is a serviceable semantic-similarity measure for retrieval and gating
Standard field practice (§3.1, §4.4); the authors note larger models might shift the optimal α (§6.4.5).

invented entities (1)

Long-term contribution accumulator ℓ_i (with short/long memory-type tag) no independent evidence
purpose: Counts reinforced activations; when ℓ_i ≥ 1.0 (~50 activations) the memory migrates to a slower decay rate γ=0.995, intended to preserve consistently used memories across sessions (Algorithm 1, lines 10-12).
Introduced as a mechanism with no falsifiable handle outside the synthetic bench; at the tested scales it is never triggered (the ablation labels it 'irrelevant at 20 rounds'), so its claimed benefit is untested.

pith-pipeline@v1.3.0-alltime-deepseek · 9516 in / 22395 out tokens · 184996 ms · 2026-08-02T13:55:42.906457+00:00 · methodology

0 comments

read the original abstract

Large language model (LLM) agents operating over extended dialogues accumulate vast amounts of information, yet existing memory systems either retain everything indiscriminately or apply uniform forgetting heuristics that fail to distinguish relevant from irrelevant knowledge. We present CAMeR (Context-Activated Memory Reinforcement), a memory retention framework combining keyword-gated hybrid activation -- a joint symbolic (word-level Jaccard) and sub-symbolic (embedding cosine) gating mechanism -- with adaptive weight dynamics. CAMeR computes a hybrid similarity score for each memory-query pair; memories exceeding a threshold receive reinforcement while all memories undergo controlled decay. We introduce CAMeR-Bench, a 76-memory, 100-round benchmark spanning 8 topic clusters with graded activation frequency, designed to test adaptive retention where existing benchmarks (LoCoMO, LongMemEval) cannot. On CAMeR-Bench, CAMeR's keyword gate achieves a 1.6$\times$ larger retention gap between high-frequency and never-referenced memories compared to embedding-only gating (scissors gap: 0.039 vs. 0.024), while time-driven baselines (Oblivion, SuperLocalMemory) collapse to near-zero weights over 100 rounds. CAMeR's top-5 retrieval saves 83.2\% tokens versus full-context approaches (39k vs. 231k cumulative) while producing weight signals that improve retrieval precision. Through 8 ablation conditions we establish that the keyword gate -- not learnable decay -- is the primary performance driver at this scale. Our findings demonstrate that hybrid symbolic-neural gating provides a simple yet effective mechanism for adaptive memory retention in LLM agents.

Figures

Figures reproduced from arXiv: 2607.20458 by Haowen Lai.

**Figure 1.** Figure 1: CAMeR system architecture. A user query triggers (1) keyword extraction and (2) embedding computation; (3) the hybrid gate combines embedding cosine and keyword Jaccard to produce an activation score; (4) the weight updater applies decay and conditional reinforcement; (5) weight-augmented retrieval ranks memories by wi× cosine similarity. Long-term migration accumulates activation counts and switches memor… view at source ↗

**Figure 2.** Figure 2: shows per-cluster final weights for CAMeR-Full. Cluster A (37 queries) reaches 0.779 while Cluster H (0 queries) drops to 0.483—a within-system gap of 0.296. The intermediate clusters (B–G) show a general downward trend correlated with activation frequency (Spearman ρ = 0.90, p < 0.01), though not perfectly monotonic due to two factors: (1) semantic adjacency between neighboring clusters (e.g., Food and He… view at source ↗

**Figure 3.** Figure 3: Cumulative token savings by retrieval method. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Figure 4: Weight trajectories from 20-round controlled experiment. Solid: activated memories [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 4 linked inside Pith

[1]

Sukhbaatar, D

S. Sukhbaatar, D. Ju, S. Poff, S. Roller, A. Szlam, J. Weston, and A. Fan. Not all memories are created equal: Learning to forget by expiring.ICML, 2021. 13

2021
[2]

V. P. Bhardwaj. SuperLocalMemory V3.3: The living brain — biologically-inspired for- getting, cognitive quantization, and multi-channel retrieval for zero-LLM agent memory systems.arXiv:2604.04514, 2026

Pith/arXiv arXiv 2026
[3]

S. Yan, X. Yang, Z. Huang, E. Nie, Z. Ding, Z. Li, X. Ma, H. Schütze, V. Tresp, and Y. Ma. Memory-R1: Enhancing large language model agents to manage and utilize memories via reinforcement learning.arXiv:2508.19828, 2025

Pith/arXiv arXiv 2025
[4]

Oblivion: Exponential time-decay heuristic for LLM memory retention.Heuristic baseline (not a published paper), 2025

2025
[5]

Mem0: The memory layer for personalized AI.https://github.com/mem0ai/ mem0, 2024

Mem0 Inc. Mem0: The memory layer for personalized AI.https://github.com/mem0ai/ mem0, 2024

2024
[6]

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.ACL, 2024

2024
[7]

D. Wu, H. Wang, W. Yu, Y. Zhang, K.-W. Chang, and D. Yu. LongMemEval: Benchmarking chat assistants on long-term interactive memory.ICLR, 2025

2025
[8]

Reimers and I

N. Reimers and I. Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT- networks.EMNLP-IJCNLP, 2019

2019
[9]

Grootendorst

M. Grootendorst. KeyBERT: Minimal keyword extraction with BERT.https://github. com/MaartenGr/KeyBERT, 2020

2020
[10]

Salton and C

G. Salton and C. Buckley. Term-weighting approaches in automatic text retrieval.Informa- tion Processing & Management, 24(5):513–523, 1988

1988
[11]

Campos, V

R. Campos, V. Mangaravite, A. Pasquali, A. Jorge, C. Nunes, and A. Jatowt. YAKE! Keyword extraction from single documents using multiple local features.Information Sciences, 509:257–289, 2020

2020
[12]

J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein. Generative agents: Interactive simulacra of human behavior.UIST, 2023

2023
[13]

Lewis, E

P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Kuttler, M. Lewis, W. Yih, T. Rocktäschel, S. Riedel, and D. Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks.NeurIPS, 2020

2020
[14]

S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao. ReAct: Synergizing reasoning and acting in language models.ICLR, 2023

2023
[15]

Packer, V

C. Packer, V. Fang, S. G. Patil, K. Lin, S. Wooders, I. Stoica, and J. E. Gonzalez. MemGPT: Towards LLMs as operating systems.arXiv:2310.08560, 2023

Pith/arXiv arXiv 2023
[16]

W. Liu, Z. Tang, J. Li, K. Chen, and M. Zhang. MemLong: Memory-augmented retrieval for long text modeling.arXiv:2408.16967, 2024. 14

Pith/arXiv arXiv 2024

[1] [1]

Sukhbaatar, D

S. Sukhbaatar, D. Ju, S. Poff, S. Roller, A. Szlam, J. Weston, and A. Fan. Not all memories are created equal: Learning to forget by expiring.ICML, 2021. 13

2021

[2] [2]

V. P. Bhardwaj. SuperLocalMemory V3.3: The living brain — biologically-inspired for- getting, cognitive quantization, and multi-channel retrieval for zero-LLM agent memory systems.arXiv:2604.04514, 2026

Pith/arXiv arXiv 2026

[3] [3]

S. Yan, X. Yang, Z. Huang, E. Nie, Z. Ding, Z. Li, X. Ma, H. Schütze, V. Tresp, and Y. Ma. Memory-R1: Enhancing large language model agents to manage and utilize memories via reinforcement learning.arXiv:2508.19828, 2025

Pith/arXiv arXiv 2025

[4] [4]

Oblivion: Exponential time-decay heuristic for LLM memory retention.Heuristic baseline (not a published paper), 2025

2025

[5] [5]

Mem0: The memory layer for personalized AI.https://github.com/mem0ai/ mem0, 2024

Mem0 Inc. Mem0: The memory layer for personalized AI.https://github.com/mem0ai/ mem0, 2024

2024

[6] [6]

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.ACL, 2024

2024

[7] [7]

D. Wu, H. Wang, W. Yu, Y. Zhang, K.-W. Chang, and D. Yu. LongMemEval: Benchmarking chat assistants on long-term interactive memory.ICLR, 2025

2025

[8] [8]

Reimers and I

N. Reimers and I. Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT- networks.EMNLP-IJCNLP, 2019

2019

[9] [9]

Grootendorst

M. Grootendorst. KeyBERT: Minimal keyword extraction with BERT.https://github. com/MaartenGr/KeyBERT, 2020

2020

[10] [10]

Salton and C

G. Salton and C. Buckley. Term-weighting approaches in automatic text retrieval.Informa- tion Processing & Management, 24(5):513–523, 1988

1988

[11] [11]

Campos, V

R. Campos, V. Mangaravite, A. Pasquali, A. Jorge, C. Nunes, and A. Jatowt. YAKE! Keyword extraction from single documents using multiple local features.Information Sciences, 509:257–289, 2020

2020

[12] [12]

J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein. Generative agents: Interactive simulacra of human behavior.UIST, 2023

2023

[13] [13]

Lewis, E

P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Kuttler, M. Lewis, W. Yih, T. Rocktäschel, S. Riedel, and D. Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks.NeurIPS, 2020

2020

[14] [14]

S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao. ReAct: Synergizing reasoning and acting in language models.ICLR, 2023

2023

[15] [15]

Packer, V

C. Packer, V. Fang, S. G. Patil, K. Lin, S. Wooders, I. Stoica, and J. E. Gonzalez. MemGPT: Towards LLMs as operating systems.arXiv:2310.08560, 2023

Pith/arXiv arXiv 2023

[16] [16]

W. Liu, Z. Tang, J. Li, K. Chen, and M. Zhang. MemLong: Memory-augmented retrieval for long text modeling.arXiv:2408.16967, 2024. 14

Pith/arXiv arXiv 2024