acceptodds
Under review as a conference paper at ICLR 2027

Q-mem: Query-Informed Online Memory for LLM Agents

Abstract

Compact recurrent memory provides LLM agents with a bounded-state mechanism for maintaining information across long-running interactions. However, as interaction histories grow, new evidence can render existing memory associations stale, creating a fundamental tension between revising outdated information and preserving useful memory. To address these challenges, we introduce Q-mem, a query-informed online memory module for frozen decoder language models. Q-mem decouples what to correct from where to update. Specifically, a gated component of the memory query orthogonal to the incoming key refines the selection of old content to revise, while the key remains the sole update address for both correction and insertion. Independent correction and write gates, together with bounded retention, separately control revision, insertion, and persistence. A request-local temporary state further allows the full user query to condition memory access without committing the request to persistent history. Across three backbones, Q-mem achieves the strongest aggregate performance among the compared methods. These results show that query-informed online revision can improve bounded recurrent memory without updating the backbone.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, memory

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_h3G3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Q-mem is a bounded recurrent memory module for frozen decoder language models. A gated query component orthogonal to the incoming key helps select old content for correction, while the key remains the sole update address. The module also uses retention, separate correction and write gates, and a temporary state for each request. Q-mem achieves the best reported aggregate LoCoMo F1 and BLEU-1 among the compared methods across three backbones. Additional experiments examine knowledge updating, context recovery, structural components, and initialization sensitivity.

Critical strengths
  1. The method offers a coherent way to separate correction-content selection from the write address. Independent correction and write gates provide flexible control, while the temporary request state lets the current question condition memory access without changing persistent history. The equations and state lifecycle are clearly described.

  2. The experiments cover three frozen backbones, several memory baselines, knowledge updating, context recovery, structural ablations, and initialization sensitivity. Q-mem consistently improves over delta-mem in aggregate LoCoMo results, and the analyses report both useful gains and limitations of the design.

Critical weaknesses
  1. Table 2 shows only small gains from the query-dependent direction over a fixed random direction: 0.45 F1 and 0.37 BLEU-1 under a single seed. The random-direction control is also insufficiently specified, including how the vector is sampled, normalized, gated, and projected against the key. Given the small gap and lack of uncertainty estimates, the results do not firmly establish the benefit of query dependence beyond adding an extra correction direction.

  2. Resetting memory across independent training episodes is reasonable, but Q-mem targets assistants that maintain user-specific memory across sessions. Although LoCoMo contains multi-session histories, the paper does not systematically evaluate retention and knowledge revision as the same persistent state undergoes progressively more updates across sessions. It therefore remains unclear whether memory quality remains stable as interactions accumulate.

Questions for authors
  1. How is the fixed random direction constructed and applied in Table 2? Can the authors report results across multiple training seeds and random-direction draws, with uncertainty, to establish whether the query-dependent gain is reliable?

  2. How is the persistent state maintained across LoCoMo sessions and queries? Can the authors evaluate retention and knowledge revision over progressively longer interactions with the same user, continuously maintaining the persistent state across sessions?

4 comments
AuthorsOpenReview ↗

Replies to Reviewer h3G3

W1&Q1: Random-direction control and reliability of query-dependent gains

In Table 2 of the original paper, we sample an eight-dimensional standard Gaussian vector for each adapted layer and memory head at initialization, normalize it to unit ℓ₂ norm, and keep it frozen throughout training and evaluation. This vector replaces the memory query only in the additional correction term and the normal query-based memory readout is unchanged. With fixed direction rr, correction gate γt\gamma_t, normalized key ktk_t, and learned strength η\eta, the correction selector is

$ a_t=\gamma_t\odot k_t+\eta(I-k_tk_t^\top)(\gamma_t\odot r). $

We apply the gate before projection and do not renormalize afterward. Thus, the sampled vector is fixed, but its gated and projected contribution still varies with the token. The update address remains ktk_t.

To quantify variability, we retrained the fixed-random control with three direction draws for each of three training seeds (42, 43, and 44). Each draw samples a new set of per-layer, per-head vectors using a separate random generator, while the other parameter initializations are held fixed within each training seed. The same LoCoMo evaluation settings are used for all nine runs.

Table R1. Fixed-random control across training seeds and random-direction draws.

Training seedDirection seedMulti-hop F1Multi-hop B1Temporal F1Temporal B1Open-domain F1Open-domain B1Single-hop F1Single-hop B1Overall F1Overall B1
42420137.3029.1637.2933.1613.8212.7656.4348.9246.2839.76
42420237.8429.0436.7932.6313.4212.2656.0848.4246.0639.32
42420338.6929.9937.2933.0413.6812.1756.5049.0746.5639.93
43430135.6628.1538.8934.7313.7712.6256.6448.3346.4339.57
43430235.7828.4439.2634.6914.0812.8755.6847.1046.0238.96
43430336.0228.8139.5135.1613.2912.4256.9548.7446.7640.00
44440136.2327.8938.1533.8413.9312.2454.8247.3745.4038.79
44440236.0428.0537.2733.2213.7412.0454.2746.9044.8638.42
44440336.4027.7838.1033.7513.3411.8556.1648.8746.1139.55

Averaging the three direction draws within each training seed, the fixed-random control achieves 46.05 ± 0.52 F1 and 39.37 ± 0.39 B1 (mean ± sample SD across the three seed means). The pooled within-seed SDs across direction draws are 0.44 and 0.48 points, respectively.

The query-dependent result in Table 2 of the original paper (46.71 F1, 40.33 B1) exceeds the fixed-random mean for each of the three control training seeds on both metrics, and exceeds 8 of 9 individual runs on F1 and all 9 on B1. Thus, the observed aggregate gain is robust to the tested choices of random direction: it does not depend on the particular draw used in Table 2. The query-dependent model was evaluated from a single training run, so this result addresses sensitivity to the random control.

AuthorsOpenReview ↗

Replies to Reviewer h3G3: second part

W2&Q2: Retention and knowledge revision over progressively longer interactions with the same user

Thank you for distinguishing multi-session QA from evaluation with a continuously updated persistent state. In the original LoCoMo evaluation (A), we reset the state before replaying each conversation, but maintain it across the sessions within that conversation. We now also evaluate a continuous stream (B and C): we process all 10 LoCoMo conversations in their stored order, carrying the same recurrent state across conversation boundaries without carrying over the attention KV cache. Questions are answered from a snapshot taken at the end of the corresponding conversation. Evaluation questions and generated answers are not written back to the persistent state.

Retention without explicit history. In C, the answer prompt contains the instruction and question but no conversation text. Table R2 compares δ-mem and Q-mem with the no-memory backbone on the same 1,540 questions.

Table R2. LoCoMo results without explicit history.

MethodState policyMulti-hop F1Multi-hop B1Temporal F1Temporal B1Open-domain F1Open-domain B1Single-hop F1Single-hop B1Overall F1Overall B1
Qwen3-4B-Instruct backboneNone8.346.942.272.4710.079.428.006.037.005.67
δ-memCarry (C)10.188.132.572.1712.579.3111.418.949.417.40
Q-memCarry (C)11.379.212.572.4412.669.7412.179.4810.057.98

To examine performance as writes accumulate, we also report F1 for the early, middle, and late portions of the stream, ending after 70, 187, and 272 sessions, respectively. Q-mem scores 10.06, 10.69, and 9.22, compared with 6.28, 6.88, and 7.70 for the backbone on the corresponding questions. The carried state therefore remains useful without explicit history even late in the stream.

Performance under continued memory updates. A and B both expose only the current conversation’s full history in the answer prompt. Table R3 compares resetting the state before each conversation (A) with carrying it across conversations (B).

Table R3. LoCoMo results with the current conversation visible in the prompt.

MethodState policyMulti-hop F1Multi-hop B1Temporal F1Temporal B1Open-domain F1Open-domain B1Single-hop F1Single-hop B1Overall F1Overall B1
δ-memReset (A)33.4426.1235.6430.7812.5511.2452.2943.1742.8935.47
δ-memCarry (B)35.9227.0230.9727.4614.6012.8453.2647.3443.0337.32
Q-memReset (A)38.5429.5538.3534.8113.8012.5356.4049.2346.7140.33
Q-memCarry (B)37.7628.1137.8733.8213.9312.4456.4749.1146.5239.79

Under carryover, Q-mem remains close to its reset result and above δ-mem. This shows that Q-mem maintains most of its QA performance when it continues updating a state that already contains earlier conversations. This indicates that, throughout increasingly long interactions with the same user, Q-mem retains a high capacity for knowledge modification.

Thank you for the clarifications and additional experiments. The random-direction controls and continuous-state experiments help address my questions. I will maintain my original positive score of 3.

Thank you for your positive feedback and for taking the time to review our additional experiments. We appreciate your constructive comments, which have helped us strengthen the paper. We will incorporate these clarifications and results into the revised manuscript.

Reviewer_1QuAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Q-mem, a query-informed online memory module for frozen decoder LMs. It decouples what old content to correct from where to update: a gated query component orthogonal to the incoming key refines stale-content selection, while the key remains the sole write address. Independent correction/write gates and bounded retention control the memory lifecycle, with a request-local temporary state for within-turn conditioning. On LoCoMo across three backbones, Q-mem achieves the strongest F1 and BLEU-1, with additional validation on LongMemEval-S and context-recovery experiments.

Critical strengths
  1. The separation of "what to correct" from "where to write" via orthogonal projection is principled and well-motivated, directly addressing a fundamental limitation of standard delta-style updates where the key conflates content selection and write addressing. The ablation comparing query-dependent vs. random correction directions confirms that query conditioning provides meaningful signal.

  2. The evaluation is methodologically thorough — structural ablations isolating each component, initialization sensitivity analysis, and context-recovery experiments across three backbones — providing credible evidence that gains stem from the proposed mechanism.

  3. The request-local temporary state is a practical design that allows full-question-conditioned memory access without polluting persistent history, which is directly useful for multi-turn agent scenarios.

Critical weaknesses
  1. The primary evaluation relies on LoCoMo (10 conversations, 1,540 questions); including additional long-term memory benchmarks would strengthen confidence in generality.

  2. Only a rank-8 single-head configuration is evaluated; understanding how state capacity scales with conversation complexity — especially for much longer interactions — is important for assessing practical applicability.

  3. The absolute improvement margins are modest on some metrics, and Context2LoRA consistently outperforms Q-mem on open-domain questions; analyzing what structural property drives this gap and whether Q-mem can be augmented to close it would be valuable.

Questions for authors
  1. How does performance trend as conversations grow far beyond LoCoMo length? Does the rank-8 state saturate?
  2. Context2LoRA consistently outperforms Q-mem on open-domain questions — have you investigated what drives this gap?
  3. See the weaknesses above.
5 comments
AuthorsOpenReview ↗

Replies to Reviewer 1QuA

W1: Generalization beyond LoCoMo on LongMemEval-S and HotpotQA

Thank you for raising this concern. We agree that LoCoMo’s 1,540 questions come from only 10 conversations, so the question count alone does not establish generality. We therefore report the full LongMemEval-S [1] results, extending the knowledge-update analysis in the submission, and a complementary evaluation on HotpotQA [2]. Both comparisons use Qwen3-4B-Instruct and the same six methods, without benchmark-specific retraining.

Table R1. LongMemEval-S: GPT-4o judge accuracy.

MethodKnowledge updateMulti-sessionSingle-sessionTemporal reasoningOverall
BM25 RAG50.0039.8568.5936.0949.40
LLMLingua-233.3312.0337.8221.8026.00
Context2LoRA55.1315.0457.0521.0536.00
MLP Memory64.1012.7854.4921.8036.20
δ-mem65.3816.5458.9724.0639.40
Q-mem67.9516.5461.5424.8140.80

Q-mem ranks second overall among the six methods (40.80%) and achieves the highest observed knowledge-update accuracy (67.95%, versus 65.38% for δ-mem and 50.00% for BM25 RAG). BM25 leads overall (49.40%). One plausible explanation is that retrieving a small set of sessions helps filter distractors in long histories. Q-mem offers a different trade-off: it updates a fixed-size recurrent state without a separate retrieval index, while performing particularly well on questions that require using revised information. These results make Q-mem a useful option when online state maintenance and knowledge updating are priorities.

Table R2. HotpotQA distractor development split: answer exact match (EM) and F1.

MethodBridge EMBridge F1Comparison EMComparison F1Overall EMOverall F1
BM25 RAG41.1356.0440.8258.8441.0756.60
LLMLingua-230.3345.0841.2957.6932.5347.61
Context2LoRA49.4165.0158.5766.9151.2565.39
MLP Memory44.9661.6839.4857.4343.8660.82
δ-mem47.1962.4057.4365.4649.2563.01
Q-mem52.1367.8456.8967.1653.0967.70

On all 7,405 HotpotQA examples, Q-mem leads the compared methods in overall EM and F1 by 1.84 and 2.31 points over Context2LoRA, the strongest overall baseline. This result extends the evaluation beyond conversational QA to multi-document reasoning.

[1] Wu, Di, et al. "LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory." arXiv preprint arXiv:2410.10813 (2024).

[2] Yang, Zhilin, et al. "HotpotQA: A Dataset for Diverse, Explainable Multi-Hop Question Answering." EMNLP 2018, pp. 2369–2380.

AuthorsOpenReview ↗

Replies to Reviewer 1QuA: second part

W2: State-capacity scaling under continuous memory updates

Thank you for raising the question of state capacity. We extended the evaluation beyond the submitted rank-8 configuration by testing ranks 4, 8, 12, 16, and 32, keeping one memory head per adapted layer and using the same frozen Qwen3-4B-Instruct backbone. The wider configurations initialize all readout dimensions while matching the initial readout norm to rank 8.

Capacity under continued updates. In the LoCoMo carryover setting, the recurrent state is carried across all 10 conversations in their stored order, while the answer prompt contains only the current conversation’s history. By the end of the stream, the state has processed 272 sessions, or approximately 221K input tokens. The table reports F1 by stream portion and overall F1/BLEU-1 on the same 1,540 questions.

Table R3. Q-mem performance across state ranks on LoCoMo.

RankState per headEarly F1Middle F1Late F1Overall F1Overall BLEU-1
44 × 443.9944.2042.3243.5436.72
88 × 845.1747.6846.0546.5239.79
1212 × 1245.6547.6042.7145.5139.12
1616 × 1647.6649.4846.1747.9441.73
3232 × 3247.1048.0943.2446.2639.79

The early, middle, and late portions end after 70, 187, and 272 sessions, respectively. Rank 16 improves on rank 8 by 1.42 F1 and 1.94 BLEU-1 points overall, showing that additional state capacity can be useful. Rank 32 does not improve further.

AuthorsOpenReview ↗

Replies to Reviewer 1QuA: third part

Q1: Long-history performance and rank-8 saturation

We evaluated rank-8 Q-mem on all 500 LongMemEval-S items, with 99K–111K history tokens per item, roughly four times the longest individual LoCoMo conversation in our runs. For each item, we reset the recurrent state, write the complete history into it, and also make that history available in the answer prompt. We sort the items by the number of tokens written to the state, split them into four equal-sized groups, and report answer token F1.

Table R4. Rank-8 Q-mem answer token F1 by history-length quartile on LongMemEval-S.

History-length quartileHistory tokensQuestionstoken F1
Q1 (shortest)99,394–105,32512532.34
Q2105,328–106,30512532.18
Q3106,317–107,10212534.12
Q4 (longest)107,119–110,73112534.04
Overall99,394–110,73150033.17

Rank-8 F1 is higher in Q3 (34.12) than in Q1 (32.34) and Q2 (32.18), but slips slightly to 34.04 in Q4. This pattern suggests that Q-mem’s benefit to the backbone may begin to taper off at the upper end of the tested length range.

W3&Q2: Open-domain performance gap with Context2LoRA

Thank you for highlighting this gap. Context2LoRA leads on LoCoMo’s open-domain questions across all three backbones, by 3.12–3.82 F1 points. We find that this gap can be substantially narrowed without changing Q-mem’s online update rule.

Table R5. LoCoMo results on Qwen3-4B-Instruct.

MethodOpen-domain F1Open-domain BLEU-1Overall F1Overall BLEU-1
Context2LoRA17.0115.6145.0139.16
Q-mem, rank 813.8012.5346.7140.33
Q-mem, rank 8, without orthogonal projection14.3712.9345.7839.34
Q-mem, rank 1613.6712.1646.0339.53
Q-mem, rank 3216.3015.1347.4440.98

Rank 32 closes 78% of the open-domain F1 gap and 84% of the BLEU-1 gap, while improving both overall scores over rank 8 in this run. Rank 16 does not show the same benefit, so the improvement is not monotonic in rank. Removing the orthogonal projection also improves open-domain scores modestly. Of the changes we tested, widening Q-mem while keeping the projection offers the better trade-off.

A plausible structural explanation is where each method acts on attention. Our Context2LoRA baseline adapts the query, key, value, and output projections. Q-mem instead reads from a compact recurrent state and injects residuals into the query and output paths only. Direct access to the key and value paths may give Context2LoRA more ways to combine conversational cues with the pretrained knowledge needed for LoCoMo’s open-domain questions [3]. The rank sweep shows that the gap can be narrowed, but it does not isolate which attention paths account for the remaining difference. A direct test would add state-conditioned key and value residuals to Q-mem while keeping its state rank, update rule, and trainable parameter budget matched. This would test whether a broader readout can close the remaining gap without sacrificing Q-mem’s overall performance.

[3] Maharana, Adyasha, et al. "Evaluating Very Long-Term Conversational Memory of LLM Agents." ACL 2024, pp. 13851–13870.

Reviewer_uy42OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper introduces Q-mem, an online memory module for frozen LLMs. The method uses a query-informed selector to choose old content for correction while retaining the incoming key as the update address.

Critical strengths
  • The paper addresses the relevant challenge of maintaining and updating compact long-term memory for LLM agents.

  • Q-mem outperforms the evaluated baselines on overall LoCoMo scores across three backbone models.

Critical weaknesses
  • The novelty seems to be limited. The proposed framework is similar to δ-mem and the changes seem incremental. The paper does not clearly explain how the method differs from δ-mem or why these differences represent a substantial contribution.

  • The source of the improvement over δ-mem is unclear. Query-informed selection seems to be the main methodological contribution, but Table 2 shows that removing it still achieves 45.13 F1, substantially above δ-mem’s 42.89 in Table 1. This suggests that other changes account for a considerable part of the improvement. More controlled comparisons with δ-mem are needed to clarify where these gains come from.

  • The benefit of query-informed selection is not sufficiently supported. In Table 2, replacing the query-dependent direction with a fixed random direction (w/o Query-Dependent Direction) already achieves 46.26 F1 and 39.96 B1, compared with 46.71 and 40.33 for the full method. Thus, query dependence contributes only 0.45 F1 and 0.37 B1. It is unclear whether these small gains reliably support the central claim.

  • The presentation is not clear enough. Figure 1 is not explicitly referenced or explained in the main text, and symbols such as (a_t) and (u_t) are not defined where the figure appears. Similarly, “incoming key” and “memory query” are introduced without explaining their roles. This makes the motivation difficult to follow without prior familiarity with existing work. A preliminary section introducing δ-mem or standard delta-style updates would help readers understand the proposed changes.

Questions for authors
  • Could the authors include a component-by-component comparison with δ-mem under matched training and evaluation settings, particularly to explain why the variant without query-informed selection already substantially outperforms it?

  • Why does Table 1 report detailed comparisons and complete metrics only on LoCoMo, rather than also including all baselines on LongMemEval-S and HotpotQA? Should the abstract’s claim that “Across three backbones, Q-mem achieves the strongest aggregate performance among the compared methods” be explicitly qualified as applying to LoCoMo?

7 comments
AuthorsOpenReview ↗

Replies to Reviewer uy42

W1: Methodological differences from δ-mem

Thank you for raising this point. We agree that the paper should distinguish more clearly between what Q-mem inherits from δ-mem and what it changes. Both methods attach a compact recurrent state to a frozen decoder and use its readout to modify the attention query and output paths. We do not claim this interface as a contribution. Q-mem instead changes which old content is revised, how retention, correction, and insertion are controlled, and how the state is used within a request.

1) Query-informed correction separates what to revise from where to write.

In δ-mem, Eqs. (5) and (10), the memory query reads the state, but the update itself is determined by the key–value pair and the gates:

$ S_t=\operatorname{Diag}(\lambda_t)S_{t-1} +\operatorname{Diag}(\beta_t)(v_t-S_{t-1}k_t)k_t^\top, \qquad \lambda_t=\mathbf 1-\beta_t. $

In particular, the old response to be corrected is St−1ktS_{t-1}k_t. Q-mem constructs a separate selector using the component of the gated query orthogonal to the normalized key:

$ a_t=\gamma_t\odot k_t +\eta_{\ell,h}(I-k_tk_t^\top)(\gamma_t\odot q_t), \qquad M_t=\bar M_t-(\bar M_ta_t)k_t^\top +(\omega_t\odot v_t)k_t^\top. $

Holding the prior state, key, value, and controllers fixed, changing qtq_t can change which old response Q-mem subtracts. The δ-mem update does not have this direct dependence on the memory query. In Q-mem, the projection prevents the added query component from simply rescaling the key-aligned part of the selector, while the shared right factor kt⊤k_t^\top keeps both correction and insertion at the original key address. Thus, the query changes what is corrected, not where the update is applied.

2) Retention, correction, and insertion are separately controlled.

δ-mem applies the same gate βt\beta_t to the old-response subtraction and the new-value insertion, and ties retention to its complement. Q-mem instead computes three separately parameterized, token-dependent controllers:

$ \begin{aligned} d_t &= \exp\!\left[-\exp(s_D)\odot \operatorname{softplus}(W_Dh_t+b_D)\right],\\ \gamma_t &= \sigma(W_\gamma h_t+b_\gamma),\qquad \omega_t = \sigma(W_\omega h_t+b_\omega),\\ \bar M_t &= M_{t-1}\operatorname{Diag}(d_t). \end{aligned} $

This changes more than the number of gates. δ-mem’s retention scales the rows of the state in value space, and its correction uses the pre-retention response St−1ktS_{t-1}k_t. Q-mem’s retention scales the columns in address space, and its correction uses the retained state Mˉt\bar M_t. Moreover, γt\gamma_t acts on the selector before the memory lookup, whereas ωt\omega_t acts on the candidate value being inserted. Q-mem can therefore revise an old association without being forced to insert new content at the same strength.

3) Request-local recurrence separates temporary adaptation from persistent history.

Q-mem initializes a temporary state from the persistent history and lets only the current question tokens update that copy:

$ \widetilde M^{(0)}=\operatorname{copy}(M^{\mathrm{hist}}), \qquad \widetilde M^{(i)} =U_\phi(\widetilde M^{(i-1)},x_i), \quad i=1,\ldots,|Q|. $

Each question token reads before it writes, so earlier question tokens can change the memory seen by later question tokens and by the answer. Answer tokens read the resulting temporary state without writing to it. After the request, the temporary state is discarded and the persistent history remains unchanged. The distinction from δ-mem’s published formulation is not that δ-mem cannot condition on a question; it is that Q-mem explicitly defines a question-only, writable state that is never committed to persistent memory. These changes matter empirically at the same recurrent state size. Across the three rank-8 backbone comparisons in Table 1, Q-mem improves over δ-mem by 2.88–4.04 F1 points and 3.26–4.86 BLEU-1 points on LoCoMo. Table 2 further separates the effects on Qwen3-4B-Instruct: from the full model’s 46.71 F1 / 40.33 BLEU-1, removing query-informed selection reduces the scores to 45.13 / 38.59, tying the correction and write gates reduces them to 44.90 / 38.30, and disabling request-local recurrence reduces them to 44.76 / 38.24. The fixed-random control recovers much of the query-selection gain, so we distinguish the benefit of adding a correction direction from the smaller observed benefit of making that direction query-dependent. Q-mem does not propose a new way to attach memory to attention. It shows how the same compact interface can support selective revision without changing the write address, and question-conditioned adaptation without changing persistent history.

AuthorsOpenReview ↗

Replies to Reviewer uy42: second part

W2/Q1: Component-wise attribution of improvements over δ-mem

Thank you for pointing this out. We agree that the full improvement over δ-mem should not be attributed to query-informed selection alone. The 45.13 F1 variant removes the query-derived correction term but retains independent correction/write gates and request-local query recurrence, as well as Q-mem’s retention rule. It is therefore not equivalent to δ-mem. To clarify where the remaining gain comes from, we retrained three additional variants that each disable two components.

Component-wise LoCoMo comparison on Qwen3-4B-Instruct. All rows use rank-8 memory, one epoch on the same 2,219 QASPER episodes, four recent messages in the main training view, and seed 42. Evaluation uses the same history-replay protocol on 1,540 questions. Q denotes query-informed correction, G independent correction/write gates, and R request-local query recurrence. Removing Q preserves key-based correction; removing G ties the correction and write gates without changing the retention rule; removing R disables the temporary question scan during both training and evaluation. Each cell reports F1 / BLEU-1 (%).

Table R1. Single- and pairwise component ablations of Q-mem on LoCoMo with Qwen3-4B-Instruct.

MethodMulti-hopTemporalOpen-domainSingle-hopOverall
δ-mem (Table 1)33.44 / 26.1235.64 / 30.7812.55 / 11.2452.29 / 43.1742.89 / 35.47
Q-mem w/o Q and G33.77 / 25.9736.80 / 32.4513.63 / 13.1352.67 / 44.4543.47 / 36.61
Q-mem w/o Q and R33.31 / 25.3634.06 / 30.2915.12 / 13.9951.85 / 44.5342.46 / 36.15
Q-mem w/o G and R37.60 / 29.0736.92 / 33.0814.80 / 13.4855.58 / 48.7845.86 / 39.70
Q-mem w/o Q36.05 / 27.9237.48 / 33.6612.31 / 10.0254.85 / 47.3145.13 / 38.59
Q-mem w/o R36.52 / 26.0136.81 / 33.3212.62 / 10.5054.23 / 47.3844.76 / 38.24
Q-mem w/o G36.01 / 27.6037.67 / 33.7813.23 / 11.7354.25 / 46.6544.90 / 38.30
Q-mem (full)38.54 / 29.5538.35 / 34.8113.80 / 12.5356.40 / 49.2346.71 / 40.33

Why does the no-query variant still outperform δ-mem? With Q disabled, tying the correction/write gates reduces F1 from 45.13 to 43.47, leaving only a 0.58-point lead over δ-mem. Alternatively, disabling request-local recurrence reduces F1 to 42.46, eliminating the lead. BLEU-1 also falls from 38.59 to 36.61 and 36.15, respectively. These comparisons show that both mechanisms contribute even without a query-derived correction term. In particular, request-local recurrence can still adapt the temporary state to the question using the key-based update.

What does query-informed correction add? Restoring Q with both other components enabled raises F1 / BLEU-1 by 1.58 / 1.74 points. Restoring Q also improves both metrics when either the gates are tied (+1.43 / +1.69) or request-local recurrence is disabled (+2.30 / +2.09). Thus, query-informed correction provides an additional benefit in each of these controlled comparisons, but it is not the sole source of the gain over δ-mem. The effects depend on which other components are enabled, so these differences should not be summed or interpreted as a complete decomposition of the gap.

AuthorsOpenReview ↗

Replies to Reviewer uy42: third part

W3: Random-direction control and reliability of query-dependent gains

Thank you for raising this concern. We agree that the single comparison in Table 2, where full Q-mem improves over the fixed-random control by 0.45 F1 and 0.37 BLEU-1 points, does not by itself establish that the gain is reliable. We clarify how the control is constructed and report nine additional fixed-random runs to assess its sensitivity to training seeds and direction draws.

How the fixed-random control is constructed. For each adapted layer and memory head, we sample a vector with independent standard Gaussian entries, normalize it to unit L2 norm, and keep it fixed throughout training and evaluation. At each token, we replace the memory query with this vector only in the additional correction term. We first multiply it elementwise by the token’s correction gate, then project the gated vector onto the subspace orthogonal to the current normalized key. We do not renormalize after projection; the projected term is scaled by the same learned bounded strength as in full Q-mem. Thus, the sampled vector is fixed, and its gated and projected contribution can still vary across tokens. The ordinary memory readout still uses the actual query, and the key remains the sole update address.

Fixed-random results across seeds and draws. We retrained rank-8 Q-mem with the fixed-random control on Qwen3-4B-Instruct using training seeds 42, 43, and 44, with three independently seeded direction draws per training seed. Each direction seed specifies a separate set of vectors across layers and heads. All runs use the same LoCoMo questions and evaluation settings. Each cell reports F1 / BLEU-1 (%).

Table R2. Fixed-random control across training seeds and random-direction draws

VariantTraining seedDirection seedMulti-hopTemporalOpen-domainSingle-hopOverall
Fixed random42420137.30 / 29.1637.29 / 33.1613.82 / 12.7656.43 / 48.9246.28 / 39.76
Fixed random42420237.84 / 29.0436.79 / 32.6313.42 / 12.2656.08 / 48.4246.06 / 39.32
Fixed random42420338.69 / 29.9937.29 / 33.0413.68 / 12.1756.50 / 49.0746.56 / 39.93
Fixed random43430135.66 / 28.1538.89 / 34.7313.77 / 12.6256.64 / 48.3346.43 / 39.57
Fixed random43430235.78 / 28.4439.26 / 34.6914.08 / 12.8755.68 / 47.1046.02 / 38.96
Fixed random43430336.02 / 28.8139.51 / 35.1613.29 / 12.4256.95 / 48.7446.76 / 40.00
Fixed random44440136.23 / 27.8938.15 / 33.8413.93 / 12.2454.82 / 47.3745.40 / 38.79
Fixed random44440236.04 / 28.0537.27 / 33.2213.74 / 12.0454.27 / 46.9044.86 / 38.42
Fixed random44440336.40 / 27.7838.10 / 33.7513.34 / 11.8556.16 / 48.8746.11 / 39.55
Full Q-mem (Table 2)42—38.54 / 29.5538.35 / 34.8113.80 / 12.5356.40 / 49.2346.71 / 40.33

The fixed-random control averages 46.05 ± 0.52 F1 and 39.37 ± 0.39 BLEU-1. We first average the three draws within each training seed; the reported standard deviations are then computed across the three seed means. At training seed 42, full Q-mem exceeds all three random draws on both overall metrics and improves over their mean by 0.41 F1 and 0.66 BLEU-1 points. The observed gain at this seed therefore does not depend on the particular random draw reported in Table 2. These results support a modest additional benefit from conditioning the correction direction on the query at the matched training seed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers

Dear Area Chair and Reviewers,

We thank all reviewers for their careful reading and constructive feedback. The discussion centered on whether Q-mem makes a meaningful change to δ-mem, which components explain its gains, and whether the evidence extends beyond the original LoCoMo setting. We conducted additional experiments to answer these questions directly. Below we summarize the main findings.

  • The 45.13 F1 no-query variant is not δ-mem: the new controls explain its remaining advantage. We added matched pairwise ablations to identify why the no-query variant still outperforms δ-mem. These comparisons show that independent correction/write gates and request-local recurrence contribute even without query-informed selection. Restoring query-informed correction further improves both metrics in each tested component context. The results clarify that the overall advantage comes from multiple interacting components, not query dependence alone.

  • The query-dependent gain survives alternative random draws at the matched training seed. We specified the random control’s sampling, normalization, gating, and projection procedures and evaluated three direction draws for each of three training seeds. At the matched training seed, full Q-mem consistently exceeds all three random draws on both overall metrics. This supports a modest additional benefit from query conditioning that does not depend on the particular random draw originally reported.

  • The gains extend beyond LoCoMo, with a clear boundary relative to retrieval. We added complete six-method comparisons on HotpotQA and LongMemEval-S without benchmark-specific retraining. Q-mem leads the compared methods on HotpotQA overall EM and F1, extending the evidence beyond conversational memory to multi-document reasoning. On LongMemEval-S, it ranks second overall and achieves the highest observed knowledge-update accuracy.

  • Continued writes do not erase the state’s usefulness over the tested stream. We added a continuous-state stress test that carries recurrent memory across all ten LoCoMo conversations without carrying the attention KV cache or committing evaluation question–answer pairs. Q-mem preserves most of its reset-condition performance under continued updates and remains above δ-mem. When explicit history is removed, the carried state still improves over the no-memory backbone. These results provide evidence that the state remains useful as updates accumulate.

  • More capacity helps, but neither scaling nor category-level improvements are monotonic. We extended the evaluation beyond rank 8 to examine state-capacity scaling. Additional capacity improves performance under continued updates, but increasing rank does not yield uniformly better results. Separate per-conversation evaluations also show that a wider state can substantially narrow the open-domain gap to Context2LoRA while improving aggregate performance. These findings characterize both the benefits and limits of increasing capacity without changing the online update rule.

Regarding Reviewer uy42’s concern about novelty relative to δ-mem, we provided a detailed, equation-level comparison that explicitly separates the inherited attention interface from the proposed changes. The distinction lies in the mathematical structure and information dependencies of the update operator itself: holding the previous state, key, value, and controller outputs fixed, changing the memory query can change Q-mem’s next state, whereas δ-mem’s update remains unchanged. This direct query dependence cannot be introduced by adjusting δ-mem’s state rank, gate initialization, or optimization settings; it requires changing the transition itself. The matched ablations establish measurable benefits from the proposed changes, providing evidence beyond architectural description alone. We therefore contend that Q-mem introduces a distinct memory-transition and request-state design whose contribution should be assessed through these structural differences and their demonstrated effects.

We thank all reviewers again for their valuable comments. We hope that this general summary is helpful for the Area Chair’s assessment of our submission, and we sincerely appreciate your time and consideration.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.