acceptodds
Under review as a conference paper at ICLR 2027

TrawMem: Reframing Agent Memory as Task-Oriented Role-Adaptive Workspaces

Abstract

Large language model (LLM) agents rely on memory systems to leverage long-term interactions, yet existing systems typically retrieve memories based on their relevance to the current query and use them as contextual evidence. However, relevance alone does not characterize how a memory contributes to task solving: the same memory may serve different roles across tasks, such as directly supporting the answer or linking to complementary memories with weaker semantic relevance but necessary information. We therefore reframe agent memory systems as Task-Oriented Role-Adaptive Workspaces, where memories are selected and organized by their task-oriented roles and relations. Based on this view, we propose TrawMem, which organizes interaction histories into local threads and constructs task-oriented workspaces by assigning roles and jointly selecting complementary memories. Experiments on LoCoMo and MemGallery with four open- and closed-source backbones show consistent improvements over representative memory systems, with macro-F1 gains of 2.23–3.14 points. On LoCoMo, TrawMem improves the LLM-Judge score over SimpleMem by 4.07 and 4.31 percentage points on Qwen3.5-9B and DeepSeek-V4-Flash, respectively, while requiring approximately 12.8% and 5.1% fewer context tokens, respectively. Codes are available at https://anonymous.4open.science/r/TrawMem.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Long-Term Agent Memory, Structured Retrieval

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_mXTVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

TrawMem argues that relevance-based memory retrieval ignores the functional role a memory plays for a given query: answering it directly versus bridging to weakly relevant but necessary evidence. The method segments dialogue history into LLM-summarized local threads linked in a thread graph. At query time it propagates scores over a bounded neighborhood and assigns each candidate turn a soft distribution over three roles (anchor/bridge/context) through fixed feature mappings. It then selects at most B turns by greedily maximizing a four-term monotone submodular objective, which carries a 1−1/e guarantee. On LoCoMo and Mem-Gallery with four answer backbones, it reports the best macro F1/BLEU-1/J against five memory systems (LoCoMo J +2.19–6.34 over the strongest baseline). It also includes ablations, sensitivity, efficiency, and mechanism diagnostics.

Critical strengths

Broad, consistent gains with reasonable robustness checks. TrawMem has the best average in all eight backbone × benchmark settings (Tables 1–2). The advantage holds under cross-model judging (Table 4: +2.26 / +5.31 J over SimpleMem) and against two more recent systems (Table 3). It also uses fewer answer tokens than SimpleMem (~2,050 vs. ~2,351 tok/q; Table 5). As a practical retrieval pipeline, it is credible. Diagnostics beyond the leaderboard. Stage-wise gold recall (Fig. 13) and the coverage–redundancy frontier against Top-B relevance (Fig. 15) give a real view of what the selection step does.

Critical weaknesses

The titular role mechanism is heuristic and its contribution is not established. Roles come from hand-designed signals (temporal/relational cues, a query-length flag) mapped through fixed, unreported matrices (App. A.3); nothing is learned. Removing role conditioning costs only 1.31–1.82 J on LoCoMo and 1.21–1.73 on Mem-Gallery (Figs. 4–5), from single runs with no variance reported. By a rough binomial estimate, the standard error of macro-J on LoCoMo is about 1.5 pp, given OpenDomain n=96. Role distributions are close to uniform (entropy 0.75–0.90 nats vs. ln 3≈1.10; Fig. 10b). Most importantly, the motivating case does not hold up. On MultiHop, final-workspace gold recall (61.3%) is below that of the initially routed threads alone (67.1%), and the same holds for Temporal (79.6% vs. 82.1%; Fig. 13). H.1.2 shows that ρ shifts across paired queries, but not that the shift changes which evidence is retrieved.

The protocol leaves the gains unattributed. (i) The answer prompt is method-specific, with CoT and per-category rules (App. J.1, R2), and the paper does not say whether baselines used it. Table 3 reports the same method on the same backbone (Qwen3.8-27B) at F1 59.74 / 54.51 under different runners, versus 51.23 in Table 1. That pipeline effect is larger than any reported method gain. (ii) Hyperparameters appear to be selected on the test set. The Fig. 6 peaks (63.26 / 62.95 / 68.56 / 74.52) exactly match Table 1, with backbone-specific (λm, λr), and B=16 is chosen on the same LoCoMo data. No held-out split is used. (iii) There is no token-matched control, such as raw-turn RAG at ~2k tokens, even though the strongest baseline, LightMem, uses about half the context (Table 5). The full-context baseline appears only in App. H.3.1 (63.70% with GPT-5.6-Luna), which is unusually low, and it is not discussed.

Questions for authors

Were all baselines run with the same answer prompt (R2) and the same judge? Please report the baselines re-run under R2, and explain the 51.23 vs. 59.74 F1 gap for the same backbone between Tables 1 and 3.

How were λm, λr, B, γ, K, H, M_max, A, and M chosen? Please report all values, along with results from a single configuration shared across backbones and fixed on a held-out split.

Please add matched-budget controls: (a) raw-turn Naive RAG at ~2,050 tokens; (b) thread retrieval + top-B by π, with no roles and no submodular selection. Please also give seed variance or bootstrap CIs for Table 1 and the ablations.

Does the role term causally recover low-α bridge evidence? For example, report gold-evidence recall stratified by α, with and without role terms, and explain the drop in MultiHop recall after selection.

12 comments
AuthorsOpenReview ↗

Response to Reviewer mXTV (Part 1/6)

Weakness 1.a & Question 2 (first half: matrices A,MA, M): The roles are obtained by mapping manually designed signals through fixed, unreported matrices, with no learning process. What are the specific values of matrices AA and MM, and how were they determined?

Thank you for your comment and question. The complete values of AA and MM are given below; both are specified once according to the role definitions in Appendix A.3 and shared across all four backbones and both datasets, with no trainable parameters. Table R1 shows that the gains come from “which feature is assigned to which role,” rather than from the specific numerical values.

Values. The rows correspond to (anchor, bridge, context), respectively; the columns of AA correspond to the query features xq=(1,ztemp,zrel,zlen)⊤x_q=(1, z_{\mathrm{temp}}, z_{\mathrm{rel}}, z_{\mathrm{len}})^\top in Eq. (17), and the columns of M(δq)M(\delta_q) correspond to the role features fq(v)=(fqa,fqb,fqc,t(v),1−fqa,1)⊤f_q(v)=(f^a_q, f^b_q, f^c_q, t(v), 1-f^a_q, 1)^\top in Eq. (20). The temperature in Eq. (5) is TR=0.50T_R=0.50 (see Table R6 in Part 5/6 for the other hyperparameters):

M(\delta_q)=\begin{bmatrix}2.20&0&0&0&0&0.35\delta_{q,a}\\\\0&1.45+0.85\delta_{q,b}&0&0&0.15&0\\\\0&0&1.30&0.85\delta_{q,c}&0.15&0\end{bmatrix}.$$ **How they were determined.** The locations of the nonzero entries follow directly from the role definitions in Appendix A.3: in $A$, the relational cue $z_{\mathrm{rel}}$ primarily increases bridge demand, while the temporal cue $z_{\mathrm{temp}}$ primarily increases context demand; in $M$, each role reads only its own features (anchor reads $f^a_q$, bridge reads $f^b_q$, and context reads $f^c_q$ and $t(v)$), and increases with the corresponding role demand $\delta_q$. The numerical values of the nonzero entries were set manually; Table R1 tests whether the results are sensitive to these values: **Table R1: Comparison of different $A$ and $M$ settings (Qwen3.5-9B)** | Matrix setting | Specific configuration | LoCoMo $J$ | MultiHop recall (%) | MemGallery $J$ | | --- | --- | --- | --- | --- | | No Role Cond. | Disable role terms | 61.95 | 55.0 | 47.85 | | Uniform $0/1$ matrices | Keep the locations of nonzero entries and set all their values to $1$ | 62.89 | 58.5 | 48.82 | | Coefficient perturbation $\pm 25$% (10 runs) | Independently multiply each coefficient by $\mathcal{U}(0.75, 1.25)$ | 63.18 $\pm$ 0.14 | 59.1 $\pm$ 0.2 | 49.02 $\pm$ 0.15 | | Random row/column permutation (10 runs) | Shuffle the feature-to-role correspondences | 60.84 $\pm$ 0.31 | 52.5 $\pm$ 0.5 | 46.62 $\pm$ 0.35 | | Original matrices (default) | As above | 63.26 | 59.2 | 49.06 | **1) The gains come from the locations of the nonzero entries, not their values (Table R1)**: When all nonzero entries are set to $1$ (with no tunable coefficients), $J$ is $62.89$ / $48.82$, already capturing $+0.94$ / $+0.97$ of the original matrices’ gains over `No Role Cond.` ($+1.31$ / $+1.21$). After randomly perturbing all coefficients by $\pm 25$% over $10$ runs, $J$ is $63.18 \pm 0.14$ (range $62.94\text{–}63.35$, with only small fluctuations around the default value). Conversely, randomly permuting rows and columns to disrupt the feature-to-role correspondence yields a result below `No Role Cond.` ($60.84$ vs. 61.95), showing that this correspondence is what matters. **2) Roles produced by the fixed mapping agree with human annotations and approach the results of per-item LLM judgments**: On $450$ samples independently annotated by two annotators ($\kappa=0.81$), the fixed mapping achieves a macro-averaged role F1 of $76.4$%, $23.9$ percentage points higher than the baseline that assigns roles by binning only $\alpha_q(v)$ ($52.5$%), and close to the $79.8$% achieved by zero-shot, per-item judgments from GPT-5.6-Luna. The latter requires one LLM call for each of approximately $145$ candidates per question, whereas the fixed mapping makes no LLM calls at query time (the F1 scores for anchor / bridge / context are $83.1$% / $71.1$% / $74.9$%, respectively). **3) The absence of training is consistent with all comparison baselines**: LoCoMo and MemGallery provide neither training sets nor role annotations; baselines such as `LightMem`, `SimpleMem`, and `Mem0` likewise have no trainable parameters during retrieval and selection. The revision will add the values of $A$ and $M$ and Table R1 to Appendices A.3 and H.1.
AuthorsOpenReview ↗

Response to Reviewer mXTV (Part 2/6)

Weakness 1.b & Question 3 (second half: variance across multiple runs): Removing role conditioning reduces JJ by only 1.21–1.82 points (close to a rough binomial standard error), and multiple-run seed variance is missing for the main results in Table 1 and all ablation variants

Thank you for your comment and question. We report the per-category results for the role ablation (Table R2) and seed variance for the original main results in Table 1 and all ablation variants in Figure 4 (Table R3), covering both paired sampling error at the question and conversation levels and the variability across 55 independent end-to-end runs (with different random seeds for Stage 1 thread induction, answer generation, and the evaluator; note that we report the mean ±\pm standard deviation over 55 independent seeds here, so normal seed-level differences of ±0.02–0.05\pm 0.02\text{–}0.05 from the single default runs in the original main table and Figure 4, such as 63.26/74.5263.26 / 74.52, are expected):

Table R2: Per-category comparison of the full model and No Role Cond. over 55 independent end-to-end runs (mean ±\pm standard deviation)

Dataset and evaluation scopeFull model (JJ)No Role Cond. (JJ)Difference ΔJ\Delta J
LoCoMo - MultiHop (n=282n=282, Qwen3.5-9B)54.68 ±\pm 0.3951.84 ±\pm 0.30+2.84±0.25+2.84 \pm 0.25
LoCoMo - Temporal (n=321n=321, Qwen3.5-9B)65.36 ±\pm 0.3463.18 ±\pm 0.41+2.18±0.22+2.18 \pm 0.22
LoCoMo - OpenDomain (n=96n=96, Qwen3.5-9B)58.54 ±\pm 0.8758.33 ±\pm 0.74+0.21±0.47+0.21 \pm 0.47
LoCoMo - SingleHop (n=841n=841, Qwen3.5-9B)74.55 ±\pm 0.1974.34 ±\pm 0.21+0.21±0.13+0.21 \pm 0.13
LoCoMo - Mean excluding OpenDomain (Qwen3.5-9B)64.86 ±\pm 0.1863.12 ±\pm 0.19+1.74±0.15+1.74 \pm 0.15
LoCoMo - Macro-JJ over all four categories (Qwen3.5-9B)63.28 ±\pm 0.2161.92 ±\pm 0.19+1.36±0.17+1.36 \pm 0.17
LoCoMo - Macro-JJ over all four categories (GPT-5.6-Luna)74.47 ±\pm 0.2372.66 ±\pm 0.25+1.81±0.19+1.81 \pm 0.19
MemGallery - Overall Macro-JJ (Qwen3.5-9B)49.09 ±\pm 0.2447.89 ±\pm 0.21+1.20±0.17+1.20 \pm 0.17

Table R3: Comparison of the original main results in Table 1 and all ablations in Figure 4 over 55 independent runs (LoCoMo, mean ±\pm standard deviation)

Comparison / ablation variantControl or variant JJFull TrawMem model JJDifference ΔJ\Delta J
Original Table 1: Qwen3.5-9B (vs. LightMem)60.01 ±\pm 0.2463.28 ±\pm 0.21+3.27±0.19+3.27 \pm 0.19
Original Table 1: Qwen3.8-27B (vs. SimpleMem)60.79 ±\pm 0.1762.91 ±\pm 0.22+2.12±0.18+2.12 \pm 0.18
Original Table 1: DeepSeek-V4-Flash (vs. LightMem)65.49 ±\pm 0.2568.59 ±\pm 0.19+3.10±0.21+3.10 \pm 0.21
Original Table 1: GPT-5.6-Luna (vs. SimpleMem)68.22 ±\pm 0.1974.47 ±\pm 0.23+6.25±0.20+6.25 \pm 0.20
Ablation (Qwen3.5-9B): Fixed Seg.61.91 ±\pm 0.2463.28 ±\pm 0.21−1.37±0.19-1.37 \pm 0.19
Ablation (Qwen3.5-9B): No Cross-Edge61.71 ±\pm 0.2363.28 ±\pm 0.21−1.57±0.20-1.57 \pm 0.20
Ablation (Qwen3.5-9B): Seed-only62.48 ±\pm 0.1663.28 ±\pm 0.21−0.80±0.15-0.80 \pm 0.15
Ablation (Qwen3.5-9B): No Role Cond.61.92 ±\pm 0.1963.28 ±\pm 0.21−1.36±0.17-1.36 \pm 0.17
Ablation (Qwen3.5-9B): Identity K62.19 ±\pm 0.2263.28 ±\pm 0.21−1.09±0.16-1.09 \pm 0.16
Ablation (Qwen3.5-9B): No Multi-thread61.37 ±\pm 0.2663.28 ±\pm 0.21−1.91±0.22-1.91 \pm 0.22
AuthorsOpenReview ↗

Response to Reviewer mXTV (Part 2/6)

1) Across question types, variation across random seeds is far below the 1.51.5 points from an independent-binomial estimate (Tables R2–R3): The 1.51.5 points correspond to two independent samples. For paired comparisons on the same 1,5401{,}540 questions, the four backbones had, respectively, 112/67112/67, 96/5996/59, 116/71116/71, and 152/68152/68 questions that changed from incorrect to correct / correct to incorrect relative to their respective strongest baselines (net gains of +37–+84+37\text{–}+84 questions, with a per-question paired standard error of only 0.81–0.960.81\text{–}0.96 percentage points). Relative to No Role Cond., about 9696%–9797% of questions had the same outcome (incorrect-to-correct / correct-to-incorrect counts ranged from 28/1128/11 to 44/1844/18 questions, with a paired per-question Macro-JJ standard error of only 0.56–0.630.56\text{–}0.63 points). Moreover, when aggregating each of the 1010 long conversations as a whole, the full model performs better on 9/109/10 conversations (two-sided sign test p=0.0215p=0.0215). Meanwhile, the standard deviation of ΔJ\Delta J across the 55 end-to-end runs is only 0.15–0.220.15\text{–}0.22 points for Table 1 and all 66 ablation variants, and the ranking of ablation importance exactly matches Figure 4.

2) Roles are especially indispensable for multi-hop, temporal, and cross-task cases (Table R2): The event-thread graph retrieves the candidate pool across sessions (hence the largest drop for No Multi-thread; we will clarify the division of responsibilities in Section 4.2), while the role module retains weakly related but complementary evidence. SingleHop and OpenDomain (n=96n=96) can be answered by direct matching (only +0.21+0.21 points; after excluding OpenDomain, the mean for the other three categories instead rises from +1.36+1.36 to +1.74±0.15+1.74 \pm 0.15), whereas roles improve JJ by +2.84+2.84 and +2.18+2.18 points on MultiHop and Temporal, respectively. On the 99 question categories in MemGallery (Qwen3.5-9B), No Role Cond. (47.89±0.2147.89 \pm 0.21) falls below the strongest baseline, Mem0 (48.51±0.2248.51 \pm 0.22); adding role modulation raises it to 49.09±0.2449.09 \pm 0.24 (outperforming in all 55 runs, ΔJ=+0.58±0.19\Delta J = +0.58 \pm 0.19).

The revision will briefly summarize these multi-run stability and category-level findings in Sections 4.1–4.2, and add the complete standard deviations, per-category role ablations, and paired statistics (Tables R2–R3) to Appendices E and H.1.

Reviewer_e4ZqOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TrawMem, a memory system for long-horizon dialogue. It segments the interaction history into local event threads that retain the raw turns, and it links threads that are adjacent in the original order or semantically similar. At query time, TrawMem locates the relevant threads and their neighborhoods and assigns each candidate memory soft weights over three task-oriented roles. It then greedily selects a complementary subset under a submodular coverage objective, and this subset forms the workspace used for answering. Experiments cover LoCoMo and MemGallery with four backbones and include ablations, an efficiency study, and mechanism analyses.

Critical strengths
  1. Well-motivated problem. Answering questions over long dialogues often requires combining scattered evidence. Selecting context by per-memory query similarity can therefore miss necessary links. Organizing evidence by its function in the current task is a reasonable direction.

  2. Practical memory design. Thread summaries serve as retrieval entry points, while raw turns, provenance, and order are preserved, so the evidence stays traceable. Combining neighborhood expansion with set-level selection is a sensible way to surface and retain complementary information.

  3. Consistent gains. TrawMem obtains the best average scores on both datasets with all four backbones. On LoCoMo, it improves over the best baseline for each metric by 2.23–3.14 F1 and 2.19–6.34 J. On MemGallery, the F1 gains range from 0.30 to 5.45.

  4. Extensive appendix analyses. The appendix includes six ablations, evaluation with a cross-model judge, stage-wise gold-source recall, a partial paired bootstrap, and an efficiency study covering context length, answer quality, and retrieval cost.

Critical weaknesses

W1. The role mechanism is not specified operationally. In Appendix A.3–A.4, the mappings A, M(δ_q), and F_a, F_b, F_c, as well as the gating g_q(u, v) in the coverage kernel, are introduced without concrete forms or a procedure for setting them. The extraction of z_temp and z_rel and the coverage threshold used for stopping (Sec. 3.2.3, Algorithm 1) are also unspecified.

Beyond hindering reproduction, this makes it hard to judge whether the computed weights reflect what a memory actually does for the query. Consider a history containing "My lease number is R17", "R17 prohibits cats", and "R18 allows cats". For "Can I keep a cat under my lease?", the first turn is the necessary bridge, while the third is only topically related. The paper does not explain how the features in Eq. (19) and the subsequent mappings separate these two cases.

W2. The independent contribution of roles is not directly established. The No Role Cond. ablation shows that the role component helps, but the paper does not state which computations it disables. Roles enter both the path cost in Eq. (6) and the role-coverage term in Eq. (8), so setting λ_r = 0 alone would not remove them. The mechanism analyses also stop short of attribution. Figure 11 shows that role weights change across queries, but query dependence alone does not show that roles track actual support or bridging.

Moreover, the 356 query pairs were generated by GPT-5.6-Luna, with no mention of human verification. Figure 15 shows better coverage and less redundancy than Top-B Relevance, and Figure 13 shows MultiHop gold-source recall falling from 79.9% after expansion to 61.3% in the final workspace. Some loss is expected under a budget, but neither analysis includes a role-ablated variant, so it remains unclear whether roles reduce the loss of key evidence.

W3. Configuration selection and evaluation protocol are under-specified, and key differences lack uncertainty estimates. The paper does not state how the main-table configuration was chosen, whether a separate development set was used, or whether baselines received comparable tuning. Notably, the grid optima in Figure 6 coincide with the Table 1 results, and the optimal setting differs across backbones. Scores also shift substantially with the runner: for GPT-5.6-Luna, TrawMem's F1 is 48.66 in Table 1 but 63.34 and 61.29 under the two runners in Table 3. These numbers are not comparable across protocols, but they call for a description of how the runners differ in prompting, history handling, budget, output parsing, and metric implementation. Finally, paired bootstrap intervals are reported only against FullText and NaiveRAG, not against the strongest baselines or for the core role ablation.

Questions for authors
  1. (W1) Please give the concrete rules for the role computation and the stopping criterion, explain how they were set, and use the lease example to show how the role scores separate the necessary bridge from the topically related turn.
  2. (W2) What exactly does No Role Cond. disable? Could you add a well-defined role-ablated variant to the analyses in Figures 13 and 15, with the same candidate set, budget, and answer model?
  3. (W3) How was the main configuration selected, and were baselines tuned comparably? Which parts of the Table 1 and Table 3 runners are shared?
  4. Table 5 shows construction time changing with H (96.1, 75.8, and 63.5 s/root for H = 3, 2, 1 on Qwen3.5-9B), although H is used only at query time (Appendix A.2, Algorithm 1). Was memory rebuilt for each configuration, or do other settings differ?
  5. MemGallery already includes KR (Knowledge Resolution) and CD (Conflict Detection). Could you report their per-category results in a table? Figure 7 shows radar charts for only three backbones. Optionally, the knowledge-update category of LongMemEval [1] and FactConsolidation-MH in MemoryAgentBench [2] could serve as external checks of reasoning over updated facts.

References:

[1] LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. ICLR 2025.

[2] Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions. ICLR 2026.

Overall Assessment

The paper offers a practical design for memory organization and evidence selection in long-horizon dialogue, and its gains are consistent across settings. I currently lean slightly toward rejection for three reasons: the role mechanism is incompletely specified, its independent contribution is not directly shown, and the configuration selection and evaluation protocol are unclear. If the authors provide reproducible role computation rules, a controlled analysis of how roles affect the retention of key evidence, and a clarification of configuration selection and protocol, I would raise my score to weak accept.

13 comments
AuthorsOpenReview ↗

Response to Reviewer e4Zq (Part 1/7)

Weakness 1 & Question 1 (first half): Provide the specific rules and rationale for the role-computation mappings (A,M,Fa,Fb,FcA, M, F_a, F_b, F_c), gate gq(u,v)g_q(u,v), query-cue extraction, and stopping criterion

Thank you for the detailed comments. This response provides complete, directly reproducible computation rules and the rationale for each coefficient. The next response (Part 2/7) walks through the full distinction process using the lease example you proposed.

(1) Specific rules for role computation, rationale, and stopping conditions (the same fixed constants are shared by all four backbones and both datasets):

  • Query cues and matrices A,M(δq)A, M(\delta_q): In Eq. (17), ztemp∈{0,1}z_{\mathrm{temp}} \in \lbrace 0,1\rbrace is triggered by case-insensitive matches to fixed temporal phrases (sentence-initial question phrases {when, what time, how long, for how long} or word-boundary phrases {what date, what year, which year, what month, what day, how many years, first time, last time, years ago, months ago, days ago}). zrel∈{0,1}z_{\mathrm{rel}} \in \lbrace 0,1\rbrace is triggered by the union of two classes of standard general lexical cues: (i) word-boundary matches to the closed set of relation-contrast and conditional prepositions {both, each other, between, in common, relationship, compare, difference, under, besides, apart from}; and (ii) question forms for counting across events or summarizing experiences (counting questions containing how many, or present-perfect questions with fronted auxiliary have/has). The compound-query length threshold is fixed at Lcomp=15L_{\mathrm{comp}}=15, i.e., zlen=1[wc(q)≥15]z_{\mathrm{len}}=\mathbf{1}[\mathrm{wc}(q)\ge 15] (the complete regular expressions will be listed in Appendix A.3 of the revision). Query demand δq=Axq/1⊤Axq\delta_q = Ax_q / \mathbf{1}^\top Ax_q in Eq. (18) is L1-normalized. The candidate role distribution ρq(v)=softmax(M(δq)fq(v)/TR)\rho_q(v) = \mathrm{softmax}(M(\delta_q)f_q(v)/T_R) in Eq. (5) uses temperature TR=0.50T_R=0.50. The 3×43\times 4 matrix AA and the 3×63\times 6 matrix M(δq)M(\delta_q) applied to the six-dimensional feature vector fq(v)=(fqa,fqb,fqc,t(v),1−fqa,1)⊤∈R6f_q(v)=(f^a_q, f^b_q, f^c_q, t(v), 1-f^a_q, 1)^\top \in \mathbb{R}^6 in Eq. (20) are:
M(\delta_q)=\begin{bmatrix}2.20&0&0&0&0&0.35\delta_{q,a}\\\\0&1.45+0.85\delta_{q,b}&0&0&0.15&0\\\\0&0&1.30&0.85\delta_{q,c}&0.15&0\end{bmatrix}.$$ - **Feature mappings** **$F_a, F_b, F_c$ (Eq. 19) and rationale**: For a candidate memory $v \in L_j$, $f^a_q(v)=0.75\alpha_q(v)+0.25r_q(T_j)$, $f^b_q(v)=0.70g_q(T_j)+0.20\partial(v)+0.10(1-f^a_q(v))$, and $f^c_q(v)=0.65\nu(v)+0.15t(v)+0.20(1-f^a_q(v))$. The inputs are taken directly from quantities already computed during retrieval: affinity $\alpha_q$, thread score $r_q$, neighborhood connectivity strength $g_q(T_j)$, thread boundary indicator $\partial$, adjacent-turn continuity $\nu$, and temporal marker $t(v)$. The coefficients are fixed according to the correspondence between roles and features (each set sums to $1$, with the main weight of $0.65\text{–}0.75$ assigned to each role’s core signal: Anchor uses query affinity $\alpha_q$, Bridge uses cross-thread connectivity $g_q(T_j)$, and Context uses within-thread continuity $\nu(v)$), without fine-grained tuning. When all coefficients in $A, M$ are randomly perturbed by $\pm 25$%, $J$ remains stable at $63.18 \pm 0.14$ over $10$ runs on Qwen3.5-9B (range $62.94\text{–}63.35$, fluctuating only slightly around the rounded default constants). - **Path activation, coverage-kernel parameters, and stopping condition**: The numerical lower bound on edge cost in Eq. (21) is $\varepsilon=10^{-9}$, and the priority-aggregation temperature in Eq. (7) is $\tau=0.30$ (per-item cap $\kappa_q=1/B_q$). The coverage kernel in Eq. (22) uses a fixed within-thread decay $\beta_{\mathrm{in}}=0.75$ and cross-thread coefficient $\beta_{\mathrm{cross}}=0.35$, with the convex-interpolation gate $g_q(u,v)=\big[(1-\eta_r)+\eta_r\sum_{r\in\lbrace a,b,c\rbrace}\delta_{q,r}\min(\rho_{q,r}(u),\rho_{q,r}(v))\big]\cdot o_q(u,v)$ (default $\eta_r=0.10$; for non-temporal queries, $o_q(u,v)=1$; for temporal queries, $o_q(u,v)=0.25+0.75\max\big(\cos_+(e_u,e_v), 0.35\max(t(u),t(v))\big)$). Algorithm 1 stops when the information-coverage term $F_q(S)$ reaches $0.70$ or the budget $B=16$ is full (to keep the context consistently around $2\text{k}$ tokens).
AuthorsOpenReview ↗

Response to Reviewer e4Zq (Part 2/7)

Weakness 1 & Question 1 (second half): Step through the lease example to show how role scores and submodular selection distinguish a “necessary bridge” memory from one that is “only topically related”

(2) Step-by-step distinction in the lease example: For the query q=q= “Can I keep a cat under my lease?”, the constraint word under matches the rule (8 words total, 8<158<15), giving xq=(1,0,1,0)⊤x_q=(1,0,1,0)^\top. L1 normalization in Eq. (18) gives δq=(0.39,0.43,0.18)\delta_q=(0.39, 0.43, 0.18) (highest demand is for Bridge). The retrieval intermediates for the three candidate memories v1,v2,v3v_1, v_2, v_3 are as follows: thread scores rq(Tj)=0.80/0.76/0.56r_q(T_j)=0.80 / 0.76 / 0.56, boundary indicators ∂(v)=1/1/0\partial(v)=1 / 1 / 0, within-thread continuity ν(v)=0.40/0.40/0.35\nu(v)=0.40 / 0.40 / 0.35, and temporal term t(v)=0t(v)=0. At τ=0.30\tau=0.30, all untruncated activation weights are below κq=1/16\kappa_q=1/16; after normalization by the total neighborhood activation mass Zq=0.362Z_q=0.362 under Eq. (7), πq(v)=0.09/0.13/0.08\pi_q(v)=0.09 / 0.13 / 0.08. The sums of candidate weights in their respective threads are 0.25/0.28/0.160.25 / 0.28 / 0.16, and the inter-thread edge weights are wH(T1,T2)=0.80,wH(T2,T3)=0.20,wH(T1,T3)=0.05w_H(T_1,T_2)=0.80, w_H(T_2,T_3)=0.20, w_H(T_1,T_3)=0.05. After substituting λm=1.25,λr=0.50\lambda_m=1.25, \lambda_r=0.50, the features, composite path cost CqC_q in Eq. (6), and marginal gains for the submodular selection objective in Eqs. (8)–(9) (comparing No Role Cond. against the full model) are shown in Table R1:

Table R1: Features, path cost CqC_q, and submodular marginal gains for the three candidate memories in the lease example (δq=(0.39,0.43,0.18)\delta_q=(0.39,0.43,0.18))

Candidate memoryαq/dq\alpha_q / d_qgq(Tj)g_q(T_j)fqa/fqb/fqcf^a_q / f^b_q / f^c_qρa/ρb/ρc\rho_a / \rho_b / \rho_cϕq\phi_q / CqC_qEq. (9) marginal gain (without role →\to full)
v1v_1: “My lease number is R17”0.56 / 0.111.000.62 / 0.94 / 0.340.35 / 0.60 / 0.050.41 / 0.960.88→0.88 \to 1.021.02 (reverses selection in favor of Bridge)
v2v_2: “R17 prohibits cats”0.72 / 0.180.650.73 / 0.68 / 0.310.68 / 0.27 / 0.050.39 / 0.831.16→1.551.16 \to 1.55 (selected in the first step as S1S_1)
v3v_3: “R18 allows cats”0.70 / 0.220.050.67 / 0.07 / 0.290.87 / 0.05 / 0.080.38 / 0.910.91→0.980.91 \to 0.98 (incorrectly selected without roles, discarded with roles)

First, based only on the single-item path cost CqC_q or non-role coverage terms, the distractor v3v_3 with high lexical overlap can outrank the bridge v1v_1: Because v3v_3 explicitly contains “allows cats,” its query relevance αq(v3)=0.70\alpha_q(v_3)=0.70 is substantially higher than that of bridge v1v_1 (0.560.56), which does not contain “cat.” Thus, even with the role term included in Eq. (6) (λr=0.50\lambda_r=0.50), the single-item path cost for v3v_3 (Cq=0.91C_q=0.91) remains lower than for v1v_1 (0.960.96). After the first-step selection S1={v2}S_1=\lbrace v_2\rbrace, if the role-coverage term is removed (No Role Cond.), v3v_3 also outranks v1v_1 in the sum of the three non-role terms (αq+Dq+ΔFq=0.70+0.16+0.05=0.91\alpha_q + D_q + \Delta F_q = 0.70 + 0.16 + 0.05 = 0.91 versus 0.56+0.25+0.07=0.880.56 + 0.25 + 0.07 = 0.88), thanks to its higher αq\alpha_q (+0.14+0.14), and is mistakenly selected first by the greedy algorithm.

Second, the cross-thread connectivity feature in Eq. (19) gives v1v_1 a high Bridge distribution, and the set-level role-coverage gain ΔRq\Delta R_q in Eq. (8) reverses the second-step ranking: Because thread T1T_1 containing v1v_1 has a strong R17 connection to the rule thread T2T_2 (gq(T1)=1.00⇒fqb=0.94g_q(T_1)=1.00 \Rightarrow f^b_q=0.94), v1v_1 obtains ρb(v1)=0.60\rho_b(v_1)=0.60. In contrast, v3v_3, which has no R17 connection (gq(T3)=0.05g_q(T_3)=0.05), remains concentrated on Anchor (ρa(v3)=0.87,ρb=0.05\rho_a(v_3)=0.87, \rho_b=0.05). After S1={v2}S_1=\lbrace v_2\rbrace has covered most of the Anchor demand (ρa(v2)=0.68\rho_a(v_2)=0.68), v1v_1 fills the unsatisfied, highest-weight Bridge demand, yielding ΔRq(v1∣S1)=0.43×(0.60−0.27)=+0.14\Delta R_q(v_1\mid S_1)=0.43\times(0.60-0.27)=+0.14. By contrast, v3v_3, which is also Anchor like v2v_2, yields only ΔRq(v3∣S1)=0.39×(0.87−0.68)=+0.07\Delta R_q(v_3\mid S_1)=0.39\times(0.87-0.68)=+0.07. This net role-coverage advantage of +0.07+0.07 directly reverses the non-role disadvantage of −0.03-0.03, so v1v_1 wins with a total marginal gain of 1.02>0.981.02 > 0.98, selecting the complete evidence chain {v2,v1}\lbrace v_2, v_1\rbrace.

The revision will add the complete closed-form equations, parameter rationale, and step-by-step example alongside Appendix A.3–A.4 and Algorithm 1.

AuthorsOpenReview ↗

Response to Reviewer e4Zq (Part 3/7)

Weakness 2.a & Question 2 (first half): The paper does not clearly specify which computations are disabled in the No Role Cond. ablation (setting only λr=0\lambda_r=0 does not fully remove the role mechanism)

Thank you for pointing out this omission. With the candidate set Uq\mathcal{U}_q, path distance dqd_q, query affinity αq\alpha_q, and budget BqB_q held fixed, No Role Cond. skips the role computation in Eq. (5) and disables all four role terms downstream:

1) Path cost in Eq. (6): Set λr=0\lambda_r=0, removing the role term λr[1−ϕq(v)]\lambda_r[1-\phi_q(v)] and retaining only path distance and query affinity;

2) Selection objective in Eq. (8): Remove the second, role-coverage term Er∼δq[max⁡u∈Sρq,r(u)]\mathop{\mathbb{E}}\nolimits_{r\sim\delta_q}\big[\max_{u\in S}\rho_{q,r}(u)\big], retaining only the other three terms;

3) Coverage-kernel gate in Appendix Eq. (22): Set the role-modulation weight ηr=0\eta_r=0 in the convex-interpolation gate, which exactly reduces to gq(u,v)=oq(u,v)g_q(u,v)=o_q(u,v);

4) Workspace assembly (Appendix J.1, Prompt R2): Do not attach the role labels anchor / bridge / context. If the workspace St⋆S_{t^\star} selected by the full model is held fixed and only the role-label text is hidden in R2, JJ for Qwen3.5-9B and GPT-5.6-Luna is 63.0963.09 and 74.3174.31, decreases of only 0.170.17 and 0.210.21, respectively, confirming that most degradation comes from candidate selection. If only λr=0\lambda_r=0 in Eq. (6) (Figure 6), JJ for the four backbones drops by just 0.16–0.510.16\text{–}0.51 points (e.g., 62.75/74.3662.75 / 74.36), whereas the complete No Role Cond. ablation (Figure 4) drops by 1.31–1.821.31\text{–}1.82 points (61.95/72.7061.95 / 72.70). This shows that the set-level role-coverage dimension in Eq. (8) contributes the most central selection gain.

The revision will briefly clarify the scope of No Role Cond. in Section 4.2, and detail all four disabled components (along with the role-label check) in Appendix D.

Weakness 2.d & Question 2 (second half): What are the results when the explicitly defined No Role Cond. is added to the Figure 13 and Figure 15 analyses with a fixed candidate set, budget, and answer model?

Thank you for this important suggestion. Keeping exactly the same candidate set Uq\mathcal{U}_q, budget B=16B=16, and answer model (GPT-5.6-Luna) as Appendix H.2 (Figures 12–15), we added No Role Cond. to the three existing strategies. The results are shown in Table R2:

Table R2: Role-ablation comparison under the same protocol as Figures 13 / 15 of the original manuscript (fixed Uq\mathcal{U}_q, B=16B=16, GPT-5.6-Luna)

Evaluation scope (expanded-stage recall reference)Selection variantGold-evidence recall (%) ↑Number of distinct relevant facts ↑Repeated fact mentions ↓LLM-Judge (JJ) ↑
MultiHop (expanded recall 79.9%)Random36.52.812.0648.94
Top-BB Relevance51.44.084.8956.74
No Role Cond.57.14.523.5159.57
TrawMem61.35.042.7863.83
Temporal (expanded recall 90.4%)Random51.22.741.9863.55
Top-BB Relevance68.93.964.7271.96
No Role Cond.76.04.413.3976.01
TrawMem79.64.882.6678.19
Macro-average over all categories (Figures 4 / 15)Random46.82.691.9461.35
Top-BB Relevance64.53.914.6070.82
No Role Cond.69.14.323.2872.70
TrawMem72.14.752.5974.52

First, corresponding to Figure 13, roles reduce the loss of gold evidence under the budget: MultiHop has an average of 145.6145.6 candidate units, and 89.589.5% are removed when assembling the workspace (Figure 12), so recall is lower than the expanded-stage value of 79.979.9%. At the same B=16B=16, No Role Cond. achieves 57.157.1% (a loss of 22.822.8 pp), while TrawMem achieves 61.361.3% (a loss of 18.618.6 pp). For Temporal, the results are 76.076.0% vs. 79.6%; the macro-average over all categories is 69.169.1% vs. 72.1%.

Second, corresponding to Figure 15, roles increase the number of distinct relevant facts while reducing repeated mentions: Compared with No Role Cond., TrawMem increases the number of distinct relevant facts from 4.324.32 to 4.754.75 and reduces repeated fact mentions from 3.283.28 to 2.592.59 (the values for Top-B Relevance are 3.91/4.603.91 / 4.60). JJ rises from 72.7072.70 to 74.5274.52 (+1.82+1.82; MultiHop +4.26+4.26, Temporal +2.18+2.18).

The revision will mark the number of units at each stage in Appendix H.2.2 and H.2.4 (Figures 13 and 15) and add the results in Table R2.

Reviewer_iverOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TrawMem, a task-oriented memory framework for LLM agents. Instead of retrieving memories solely by semantic relevance, TrawMem organizes interaction histories into event threads and assigns task-dependent roles (anchor, bridge, and context) to select complementary memories. Experiments on LoCoMo and MemGallery demonstrate consistent improvements across four LLM backbones.

Critical strengths
  • The paper highlights the importance of memory roles and relationships beyond semantic relevance, providing an interesting perspective on agent memory retrieval.
  • The proposed framework effectively combines thread-based memory organization, role assignment, and submodular selection.
  • The experiments demonstrate consistent improvements across multiple benchmarks and LLM backbones, supported by extensive ablation studies.
Critical weaknesses
  • The method combines several existing techniques, and the unique contribution of role adaptation is not sufficiently isolated.
  • The role assignment mechanism relies heavily on handcrafted features without direct evaluation of role assignment accuracy.
  • The proposed framework introduces additional computational overhead despite reducing answer-context tokens.
  • The evaluation focuses on two benchmarks without sufficiently demonstrating scalability to larger, dynamically evolving memory systems.
Questions for authors
  • Can the authors better isolate the benefits of role adaptation from graph expansion and diversity-aware retrieval?
  • How robust is the heuristic role assignment mechanism to different query formulations and memory structures?
  • Why does the final workspace selection substantially reduce gold-source recall for multi-hop questions (Figure 13)?
  • How does the proposed framework scale to significantly larger memory repositories, particularly in terms of retrieval latency and computational cost?
7 comments
AuthorsOpenReview ↗

Response to Reviewer iver (Part 1/3)

Weakness 1 & Question 1: How can the unique benefits of role adaptation be more clearly isolated from graph expansion and diversity-aware retrieval?

Thank you for your comment and question. Standard graph expansion and diversity retrieval (e.g., MMR) consider only whether semantics differ; they cannot distinguish a key memory with low lexical overlap that serves as a bridge from an irrelevant dialogue. To independently validate the role of role adaptation, we performed a stepwise ablation of graph expansion, diversity/coverage selection, and role adaptation at a fixed budget of B=16B=16 (Table R1):

Table R1: Stepwise comparison of graph expansion, diversity/coverage selection, and role adaptation (LoCoMo, budget B=16B=16)

Comparison variant (all use budget B=16B=16)Graph expansionDiversity/coverage selectionRole adaptationQwen3.5-9B JJGPT-5.6-Luna JJMultiHop recall (%, GPT-5.6)
1. Seed threads + Top-BB relevance (by αq\alpha_q)✗✗✗58.6269.4550.2
2. Graph expansion + Top-BB relevance (by αq\alpha_q)✓✗✗59.8470.8251.4
3. Graph expansion + classic diversity retrieval (MMR)✓MMR✗61.1271.8854.6
4. Graph expansion + role-free submodular coverage (No Role Cond.)✓Submodular kernel KqK_q✗61.9572.7057.1
5. Graph expansion + role adaptation (without kernel coverage, Identity K)✓✗✓62.1573.0558.6
6. TrawMem (full model)✓Submodular kernel KqK_q✓63.2674.5261.3

1) Standard “graph expansion + diversity retrieval” cannot replace role adaptation (variants 3 and 4 vs. 6): With the same expanded candidate set Uq\mathcal{U}_q, using classic diversity retrieval MMR (variant 3) gives JJ scores 2.142.14 points (61.1261.12 vs. 63.2663.26) and 2.642.64 points (71.8871.88 vs. 74.5274.52) below the full model on the two backbones, respectively. MultiHop recall is lower by 6.76.7 percentage points (54.654.6% vs. 61.361.3%). Even when the full submodular coverage kernel is retained and only roles are disabled (variant 4, No Role Cond.), JJ still drops by 1.311.31 and 1.821.82 points, and MultiHop recall drops by 4.24.2 percentage points (recall of low-αq\alpha_q bridging evidence falls from 41.141.1% to 33.633.6%). This is because standard diversity only encourages non-redundant content and can select off-topic dialogue, whereas role adaptation specifically fills in bridging and contextual information.

2) Role adaptation alone and the submodular coverage kernel alone make comparable, slightly different contributions, and are complementary when combined (variants 4 and 5 vs. 6): With graph expansion enabled, “role adaptation only, without the coverage kernel” (variant 5, Identity K; 62.15/73.0562.15 / 73.05, MultiHop recall 58.658.6%) and “coverage kernel only, without role adaptation” (variant 4, No Role Cond.; 61.95/72.7061.95 / 72.70, recall 57.157.1%) both substantially improve over pure relevance ranking (variant 2; 59.84/70.8259.84 / 70.82, recall 51.451.4%), with role adaptation alone performing slightly better. Combining the two (variant 6; 63.26/74.5263.26 / 74.52, recall 61.361.3%) yields the largest gain, showing that role-based completion and coverage-based deduplication are complementary rather than interchangeable.

3) Role adaptation preserves bridging clues retrieved by graph expansion for multi-hop reasoning; for overall scores, the two have partially overlapping functions as well as independent gains: For overall JJ across the four categories, the initial K=16K=16 seed threads already contain some local connections, so the marginal gain from graph expansion in the full model (from Seed-only to the full model in original Figure 4: +0.81/+1.22+0.81 / +1.22 points) is naturally a little smaller than its gain over the pure relevance Top-B baseline (variant 1 →\to 2: +1.22/+1.37+1.22 / +1.37 points), reflecting partial overlap in their functions. However, for cross-session bridging measured by MultiHop recall, graph expansion alone without roles (variant 1 →\to 2) improves recall by only +1.2+1.2 percentage points (50.250.2% →\to 51.451.4%), because low-αq\alpha_q bridging memories retrieved through expansion are discarded when truncated by relevance. After adding submodular coverage and role adaptation, graph expansion raises MultiHop recall to 61.361.3% relative to the same-budget Seed-only result (53.853.8%), a gain of +7.5+7.5 percentage points, corresponding to MultiHop ΔJ\Delta J of +1.77/+2.84+1.77 / +2.84 points. This shows that the two stages play complementary roles in preserving multi-hop evidence.

The revision will briefly summarize the stepwise isolation of role adaptation and the MMR comparison in Section 4.2, and add the full results in Table R1 to Appendix H.2.

AuthorsOpenReview ↗

Response to Reviewer iver (Part 2/3)

Weakness 2 & Question 2: The role-assignment mechanism relies on hand-designed features and lacks direct accuracy evaluation. How robust is it to changes in query wording and memory structure?

Thank you for your comment. We directly evaluate the heuristic role-assignment mechanism in two respects: blind human evaluation of role-assignment accuracy and robustness to query wording and graph-structure perturbations.

(1) Direct evaluation of role-assignment accuracy (Table R2): We stratified and sampled 450450 (query, memory) pairs from LoCoMo examples with gold evidence. Two annotators independently labeled each memory’s actual function (agreement κ=0.81\kappa = 0.81, with disagreements adjudicated by a third annotator; human labels were not used to set parameters), and we compared these labels against the model prediction arg⁡max⁡rρq,r(v)\arg\max_r \rho_{q,r}(v):

Table R2: Accuracy comparison between model-inferred roles and human-annotated roles (LoCoMo, N=450N=450)

Role category / methodNumber of samplesPrecision (%)Recall (%)F1 (%)
Anchor (direct support)17882.084.383.1
Bridge (cross-hop bridge)13273.468.971.1
Context (background supplement)14074.175.774.9
TrawMem macro-average45076.576.376.4
Baseline assigning roles only by binning αq(v)\alpha_q(v) (macro-average)45052.952.152.5

(2) Robustness evaluation under query paraphrases and memory-graph structure perturbations (Table R3): On LoCoMo (Qwen3.5-9B, B=16B=16), we tested both meaning-preserving query paraphrases (stratified sample of 100100 questions ×4\times 4 paraphrases, manually verified with κ=0.86\kappa=0.86) and three types of memory-graph perturbations on all 1,5401{,}540 questions (unperturbed overall macro-average recall is 70.470.4%, and J=63.26J=63.26):

Table R3: Robustness to query paraphrasing and memory-structure perturbations (LoCoMo, Qwen3.5-9B)

PerturbationEvaluation sizeDominant-role consistency (%) / ρq\rho_q JS ↓Workspace Jaccard overlap ↑Macro-average gold recall (%) ↑ΔJ\Delta J vs. original
Meaning-preserving query paraphrases (xqx_q consistency 94.594.5%)100100 queries ×4\times 493.8 / 0.050.8269.9−0.25-0.25
Random deletion of 1010% of cross-thread edgesAll 1,5401{,}540 questions92.4 / 0.060.7869.1−0.38-0.38
Random deletion of 2020% of cross-thread edgesAll 1,5401{,}540 questions88.1 / 0.110.7167.4−0.91-0.91
Shift event-thread boundaries by ±1\pm 1 turnAll 1,5401{,}540 questions90.6 / 0.080.7568.7−0.52-0.52

1) Model-predicted roles closely match human judgments; structural and temporal features are important for identifying bridge roles (Table R2): Across 450450 samples, TrawMem achieves macro-average Precision / Recall / F1 of 76.576.5% / 76.376.3% / 76.476.4%, with F1 above 7171% for all three roles (Anchor reaches 83.183.1%). Compared with the binned baseline using only query relevance αq(v)\alpha_q(v) (macro F1 of 52.552.5%), adding cross-thread connectivity and within-thread continuity features increases macro F1 by +23.9+23.9 percentage points. The gain is especially large for the Bridge class with low lexical overlap, where F1 rises from 41.841.8% to 71.171.1% (+29.3+29.3 percentage points).

2) The features reuse retrieval intermediates and remain robust to coefficient perturbations, query wording, and graph-structure changes (Table R3): Each feature in Eq. (19) is taken directly from affinity and graph-connectivity scores already computed during retrieval, requiring neither role-annotation training nor additional query-time LLM calls. With ±25\pm 25% random perturbations to the mapping matrices over 1010 runs, JJ remains at 63.18±0.1463.18 \pm 0.14 (range 62.94–63.3562.94\text{–}63.35). Under meaning-preserving paraphrases, query cues xqx_q are consistent componentwise in 94.594.5% of cases (only 5.55.5% flip), with a mean L1 shift of only 0.030.03 in demand distribution δq\delta_q; the dominant candidate role arg⁡max⁡rρq,r(v)\arg\max_r \rho_{q,r}(v) is consistent in 93.893.8% of cases (JS divergence 0.050.05), workspace overlap is 0.820.82, and ΔJ\Delta J changes by only −0.25-0.25 points. Even when 2020% of cross-thread edges are randomly removed or event boundaries are shifted by ±1\pm 1 turn across all questions, ΔJ\Delta J declines gradually by only 0.910.91 and 0.520.52 points, respectively, and remains more than 2.32.3 points above the strongest baseline.

The revision will add Tables R2–R3 to Appendix H.1, including human evaluation of role-assignment accuracy, coefficient perturbations, and robustness results.

AuthorsOpenReview ↗

Response to Reviewer iver (Part 2/3)

Weakness 3: Although the framework reduces answer-context token count, it still introduces additional computation overhead overall

Thank you for this observation. Using the measured data in Appendix F (Table 5) of the original manuscript, we clarify the runtime comparison, sources of overhead, and tunability below:

1) Under the default setting, total time is below the strongest baseline LightMem and substantially faster than comparable graph-memory methods: With default H=3H=3, TrawMem’s total time is slightly below that of the highest-accuracy baseline, LightMem, on both Qwen3.5-9B (699.9 s/root699.9\text{ s/root} vs. 711.8 s711.8\text{ s}) and DeepSeek-V4-Flash (513.6 s/root513.6\text{ s/root} vs. 515.2 s515.2\text{ s}), while JJ is higher by +3.30+3.30 and +3.02+3.02 points, respectively. Compared with Mem0 and A-MEM, total time is reduced by 52.452.4%–88.488.4%. Relative only to the speed-oriented SimpleMem (605.3 s605.3\text{ s} / 441.3 s441.3\text{ s}), total time increases by about 1616%, but context is shortened by 5.15.1%–12.812.8% and JJ improves by +4.07–+4.31+4.07\text{–}+4.31 points.

2) Query-time extra computation never calls an LLM, and the per-question latency increase is small: After correcting the timing boundary between offline graph construction and query routing, query-stage time is 643.5 s643.5\text{ s} and 480.1 s/root480.1\text{ s/root} (total time is unchanged; about 154154 questions/root on average). The added time comes from multi-hop neighborhood expansion (about 0.02–0.260.02\text{–}0.26 seconds per question) and workspace selection. Role scoring (fixed matrix multiplication) and greedy selection (about 145145 candidates and at most 1616 selections) are lightweight numerical computations that add no LLM calls. Per question, query-stage time increases by only about 0.270.27 seconds over the strongest baseline, LightMem, and by 0.68–0.920.68\text{–}0.92 seconds over the fastest baseline, SimpleMem. The shorter context also directly reduces downstream LLM generation cost.

3) Reducing the number of hops HH matches the fastest baseline while retaining accuracy advantages: As shown in Appendix Table 5 of the original manuscript, the intermediate setting H=2H=2 reduces total time to 628.5 s628.5\text{ s} and 446.8 s/root446.8\text{ s/root}, matching the fastest baseline, SimpleMem (within 1.21.2% under DeepSeek), while keeping context compression unchanged and maintaining a higher JJ than SimpleMem (+2.39+2.39 and +0.77+0.77 points).

The revision will add average per-question latency in Appendix F and practical guidelines for trading off latency and accuracy.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Area Chair and All Reviewers

Thank you to the Area Chair and the three reviewers (mXTV, e4Zq, and iver) for the careful reviews and constructive feedback. We have responded point by point to every Weakness and Question raised by the three reviewers, and addressed each issue through additional experiments, clarifications, or direct revisions to the original manuscript (see the responses under each reviewer). The main additions are summarized below:

1) Complete details of role computation (mXTV, e4Zq, iver): We provide all formulas and parameters used in role computation and walk through an example proposed by a reviewer step by step. New experiments show that model-assigned roles closely match human judgments and remain stable under parameter changes, query paraphrases, and changes in memory structure.

2) Practical effect of the role mechanism (mXTV, e4Zq, iver): We specify exactly which computations are disabled by the role-free comparison setting. With the same number of memories retained, we add stepwise comparisons with several retrieval and selection strategies and an experiment that directly replaces memory roles. These results show that the role mechanism helps preserve key memories that have low literal relevance to the question but are essential for answering it.

3) Fairness of comparisons and stability of results (mXTV, e4Zq): We confirm that all main-table baselines use the same prompts and evaluation procedure, and explain the prompt and runner differences in the appendix comparisons. We add parameter selection on independent conversations, one shared parameter configuration across all models, comparisons at the same context length, and multiple independent runs. The results show that the improvements are stable and do not depend on per-model parameter tuning or longer context.

4) Runtime efficiency and larger-scale settings (e4Zq, iver): We correct the timing-boundary misalignment in Appendix Table 5 (total time remains unchanged). After correction, our method has lower memory-construction time than all baselines, and lower total time than LightMem, Mem0, and A-MEM (while matching the speed-oriented SimpleMem at H=2H=2). We add latency tests with an 88-fold larger memory store and complete results on memory-update and conflict tasks, showing that the method also applies to larger, continuously changing memory stores.

All of the above additions and revisions will be incorporated into the updated manuscript. Thank you again to the Area Chair and all reviewers for your valuable feedback!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.