acceptodds
Under review as a conference paper at ICLR 2027

DyadMem: Dual-Zone Memory with Source-Grounded Recall for Long-Horizon Agent Tasks

Abstract

Long-horizon agentic tasks produce trajectories that exceed LLM context windows, creating a compression–traceability dilemma: compact memories support global understanding but can obscure details and provenance, whereas chunk retrieval does not explicitly preserve trajectory structure. Agent histories therefore require both an up-to-date task state and a way to locate precise evidence across ordered action–observation steps. We propose DyadMem, a dual-zone memory framework comprising Situation Memory and Event List. Situation Memory maintains a cumulative summary of goals, progress, decisions, and constraints, while Event List forms a chronological index whose stable identifiers map to immutably preserved raw records. Memory construction is question-independent, and hierarchical re-compression preserves identifier coverage. At query time, the agent drafts an answer from memory and optionally verifies it through find_raw_events, which supports natural-language search, identifier-based reading, and focused re-reading. On CAME-Bench, DyadMem achieves 76.17% F1 with Gemma-4-31B and 64.29% with Qwen3-32B; on AMA-Bench, it reaches 67.75% and 66.15% accuracy, respectively, establishing new state-of-the-art results across both benchmarks and outperforming eight strong baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Agent Memory

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_6ZjzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DyadMem, a dual-zone memory framework for long-horizon LLM agents. It combines Situation Memory, which maintains a cumulative summary of task states, with an Event List, which preserves chronological event summaries linked to original interaction records. The framework additionally supports hierarchical memory compression and source-grounded retrieval through natural-language search, event-ID lookup, and focused re-reading. Experiments on CAME-Bench and AMA-Bench demonstrate consistent improvements over long-context, retrieval-based, and memory-augmented baselines across two backbone models.

Critical strengths
  1. Strong empirical results. DyadMem achieves 76.17% F1 on CAME-Bench and 67.75% accuracy on AMA-Bench with Gemma-4-31B, outperforming eight baselines. The improvements are reproduced with Qwen3-32B.

  2. Reasonable motivation and practical design. Combining cumulative task-state summaries with chronological event indexing and raw-record retrieval addresses a genuine limitation of compressed agent memory.

  3. Meaningful component ablations. Removing Situation Memory or Raw Recall reduces CAME-Bench performance by approximately 9 points, suggesting that both contribute to the framework.

Critical weaknesses

1. Limited methodological novelty.

The proposed framework primarily combines established ideas: cumulative summarization, chronological indexing, hierarchical compression, and source-grounded retrieval. Similar mechanisms have already been explored in ReadAgent, STITCH, AMA-Agent, and other agent-memory systems.

The paper does not clearly demonstrate a fundamentally new memory mechanism beyond their integration.

2. The ablations provide limited evidence for the novel design choices.

Table 3 is particularly concerning:

Design componentF1 improvement
Hierarchical compression vs. FIFO+1.09
Explicit Event IDs+1.34
Named-anchor expansion and connection hints+0.64
Proposed retrieval vs. BM25 retrieval+0.93

These improvements are relatively small. Moreover, removing Event IDs actually improves performance on the Large subset.

The substantial improvement from introducing Event List demonstrates the importance of retaining historical information, but does not establish that the specific proposed indexing design is superior to simpler, budget-matched alternatives.

3. Insufficient evidence of practical effectiveness and computational efficiency.

Although the paper targets long-horizon agent tasks, the evaluation primarily measures question-answering accuracy over recorded trajectories rather than improvements in actual agent execution.

Furthermore, DyadMem uses repeated memory construction, compression, drafting, and potentially multiple retrieval calls. The paper does not provide a systematic comparison of total token consumption, inference latency, or computational cost.

The baseline comparison also deserves closer scrutiny: the standalone RAG baselines do not use iterative retrieval, whereas DyadMem can perform multi-round recall.

Questions for authors
  1. Methodological novelty: How does DyadMem fundamentally differ from existing memory architectures such as ReadAgent, STITCH, and AMA-Agent? Could the authors compare against a simpler baseline combining chronological summarization with raw-record retrieval under an identical memory budget?

  2. Limited gains from the proposed design choices: Table 3 shows that hierarchical compression, explicit Event IDs, and the proposed retrieval refinements provide only modest improvements (approximately 0.6–1.3 F1 points). Moreover, removing Event IDs improves performance on the Large subset. How do the authors justify the necessity of these specific components? Are these improvements statistically significant?

  3. Baseline fairness and computational efficiency: DyadMem benefits from multi-round retrieval and draft-then-verify reasoning, while standalone RAG baselines use single-pass retrieval. Could the authors evaluate stronger baselines equipped with comparable iterative retrieval capabilities? Additionally, how do total token consumption, inference latency, and computational costs compare when memory construction and repeated compression are included?

  4. Generalization to real agent execution: The experiments evaluate memory-based question answering over recorded trajectories rather than actual long-horizon agent execution. Can the authors demonstrate that DyadMem improves downstream task success in interactive agent environments, rather than primarily improving retrospective information retrieval?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3 (Part 1/3)

We thank the reviewer for recognizing our empirical results, motivation, and ablation design. We have completed all the suggested experiments and respond point by point below.

W1 / Q1: Methodological novelty

We agree that cumulative summarization, chronological indexing, hierarchical compression, and raw-record retrieval each have precedents on their own. Our contribution is a dual-zone principle for organizing memory:

  1. "What is the current state" (Situation Memory) and "what happened in the past and where" (Event List) are maintained separately, each with its own budget;
  2. The Event List is a chronological index covering the entire history: every summary links to its raw records, and when the budget is exceeded, adjacent summaries are merged rather than discarded, so the history remains gap-free while staying within budget;
  3. At answer time, the model first drafts from these two zones, then goes back to the raw records to verify the specific gaps in the draft.

Compared with existing methods, the main difference lies in how memory is organized. ReadAgent generates a gist once per page and never compresses it further, so gist memory grows linearly with the trajectory and exceeds the context in our 32k-budget experiments. STITCH organizes memory around contextual intents, and AMA-Agent builds a causal graph combined with tool-based retrieval; both organize history by semantic relations. DyadMem instead organizes history by time: alongside a separately maintained current state, it keeps a budget-bounded event index that covers the entire history and links to the raw records.

We have completed the control the reviewer suggested, "chronological summarization + raw-record retrieval under an identical memory budget" (Simple; see the General Response for the setup). Simple is such a simpler alternative: a single summary for both state and history, plus a coarse-grained table of contents.

CAME/GemmaOverall F1SmallMediumLarge
Simple60.5381.3853.3031.20
DyadMem76.1785.5078.3848.08

DyadMem outperforms Simple by 15.64 F1, and by 25.08 and 16.88 on Medium and Large trajectories, respectively. This suggests that, under the same raw retrieval and memory budget, the combination of chronological summarization and raw retrieval cannot substitute for DyadMem's memory organization. Combined with the w/o Situation ablation (which keeps the per-record Event List and retrieval settings unchanged and removes only the state zone, reducing F1 by 9.01), this indicates that the state zone and the per-record history zone each contribute.

W2 / Q2: Necessity and statistical significance of the smaller design choices

Following the reviewer's suggestion, we ran a trajectory-level paired bootstrap (20,000 resamples) for each comparison in Table 3:

Full method − variantΔ F195% CI
w/o Situation+9.01[+3.01, +14.03]
w/o Raw Recall+9.24[+3.76, +14.31]
w/o Event+39.18[+23.46, +51.12]
FIFO Event List+1.09[−1.07, +3.57]
w/o Event IDs+1.34[−2.87, +5.49]
w/o Named-Anchor Expansion+0.64[−2.12, +2.99]
BM25 Retrieval+0.93[−2.10, +3.43]

The first three rows correspond to the two memory zones and raw recall, and none of their intervals include zero; the Simple control above also supports the role of the dual-zone organization. The intervals of the last four rows include zero, and the reviewer's caution is reasonable. However, these four components play different roles and are best read separately:

  • Hierarchical re-compression (vs. FIFO). Its role is mainly structural: under a limited budget, old events are merged rather than discarded, so the Event List always covers the entire history. This difference only emerges when trajectories are long enough for the Event List to repeatedly reach its capacity: on the Large subset, FIFO scores 41.35 versus 48.08 for the full method.
  • Visible IDs. w/o Event IDs removes only the IDs rendered to the model; the internal source mapping is fully preserved, and retrieval grouping, the coverage relation, and reading raw records by ID all work as usual. This ablation therefore measures the benefit of the model seeing IDs directly, not whether source mapping itself is necessary. Its higher score on the Large subset indicates that visible IDs by themselves play a limited role.
  • Named-anchor expansion and retrieval backend. These are retrieval-tool-level choices with a small effect on average F1 (+0.64, +0.93), and we will reduce their emphasis in the main text accordingly.

We will add the confidence intervals to Table 3 in the revision.

AuthorsOpenReview ↗

Response to Reviewer 3 (Part 2/3)

W3 / Q3: Iterative-retrieval baseline

We thank the reviewer for this suggestion; this baseline should indeed be included. We added an Agentic Dense RAG baseline. It uses the same index as Dense RAG, but the model retrieves autonomously via tool calls for up to 6 rounds (vs. 3 for DyadMem) and can rewrite its queries; the context is likewise 32k.

MethodCAME/Gemma F1CAME/Qwen F1AMA/Gemma AccAMA/Qwen Acc
Dense RAG36.9333.9246.4750.00
Agentic Dense RAG44.8729.3653.0448.72
DyadMem76.1764.2967.7566.15

Iterative retrieval helps on Gemma and slightly hurts on Qwen, and the gap to DyadMem remains 14.7–34.9 points. In addition, the Simple control uses the same draft-then-verify procedure and reaches only 60.53 F1. Multi-round retrieval and verification by themselves therefore do not explain the main gains.

W3 / Q3: Computational cost

We measured the token cost of generation-model calls for all 10 methods, split into memory construction (including all re-compressions) and answering. Amortized cost = (total construction cost + total answering cost) ÷ total number of questions. Each cell reads "total input (of which non-cached input) / output"; total input includes server-side prefix-cache hits and excludes embedding and judge calls. "–" means the stage does not call the generation model.

CAME/Gemma (14 trajectories, 373 questions)

MethodF1Construction (per trajectory)Answering (per question)Amortized (per question)
DyadMem76.17480.0k (443.0k) / 34.7k26.6k (5.0k) / 0.08k44.6k (21.7k) / 1.4k
Long Context (256k)72.98–142.9k (45.3k) / 0.02k142.9k (45.3k) / 0.02k
AMA-Agent37.68195.8k (192.9k) / 27.7k11.7k (9.9k) / 1.2k19.1k (17.2k) / 2.2k
Agentic Dense RAG44.87–33.2k (24.5k) / 0.25k33.2k (24.5k) / 0.25k
ReadAgent-P44.28601.5k (553.2k) / 58.9k24.8k (5.7k) / 0.23k47.3k (26.4k) / 2.4k
STITCH44.9315.2M (14.3M) / 114.3k23.6k (13.9k) / 1.2k594.2k (548.8k) / 5.5k
HippoRAG 229.39353.2k (308.7k) / 18.5k7.9k (5.3k) / 0.21k21.1k (16.9k) / 0.90k
Dense RAG36.93–4.4k (4.2k) / 0.02k4.4k (4.2k) / 0.02k
BM25 RAG23.74–4.4k (4.2k) / 0.02k4.4k (4.2k) / 0.02k
MemGPT12.44–12.8k (9.3k) / 0.07k12.8k (9.3k) / 0.07k

AMA/Gemma (208 trajectories, 2,496 questions)

MethodAccConstruction (per trajectory)Answering (per question)Amortized (per question)
DyadMem67.7575.1k (74.5k) / 4.5k31.1k (21.2k) / 0.37k37.3k (27.4k) / 0.74k
Long Context (256k)60.94–54.1k (26.8k) / 0.09k54.1k (26.8k) / 0.09k
AMA-Agent60.2250.3k (49.2k) / 6.4k9.1k (8.5k) / 0.35k13.3k (12.6k) / 0.88k
Agentic Dense RAG53.04–19.7k (12.5k) / 0.36k19.7k (12.5k) / 0.36k
ReadAgent-P49.84408.5k (394.5k) / 22.4k14.0k (6.2k) / 0.12k48.0k (39.1k) / 2.0k
STITCH30.971.22M (1.06M) / 15.9k14.1k (9.1k) / 0.70k115.6k (97.5k) / 2.0k
HippoRAG 245.43173.5k (153.3k) / 18.5k11.6k (9.3k) / 0.12k26.1k (22.0k) / 1.7k
Dense RAG46.47–4.5k (4.0k) / 0.09k4.5k (4.0k) / 0.09k
BM25 RAG41.03–4.5k (3.9k) / 0.09k4.5k (3.9k) / 0.09k
MemGPT30.21–18.3k (13.7k) / 0.15k18.3k (13.7k) / 0.15k

Latency. Concurrency and caching conditions differ across methods in the full runs, so their speeds are not directly comparable. We therefore ran a controlled experiment: on AMA/Gemma, we fixed 18 trajectories and 216 questions (6 domains × one short/medium/long trajectory each), and every method built its memory from scratch and answered questions one by one on a single instance with a single concurrent request.

MethodConstruction (s/trajectory)Answering (s/question)Amortized (s/question)
DyadMem228.523.542.5
Long Context (256k)–73.973.9
AMA-Agent236.020.740.3
Agentic Dense RAG11.431.932.8
ReadAgent-P1,444.39.6130.0
STITCH†1,327.243.2136.0
HippoRAG 2†1,045.413.498.3
Dense RAG11.47.98.9
BM25 RAG0.57.47.5
MemGPT27.911.914.2

The construction time of Dense RAG and Agentic Dense RAG is the time to build the embedding index. † These two methods failed to complete some questions on this subset. The Answering column averages over completed questions, while the Amortized column is computed over all 216 questions with no answering time counted for unfinished ones; their amortized values are therefore underestimated.

AuthorsOpenReview ↗

Response to Reviewer 3 (Part 3/3)

W3 / Q3: Computational cost (continued)

These results show three things:

  • Answering overhead is moderate. The constructed memory is shared across questions and reused as a cached prefix at answer time. On CAME, only 5.0k of DyadMem's 26.6k answering input per question is non-cached, comparable to Dense RAG (4.2k), and 57% of the questions invoke no retrieval at all. In the controlled experiment, DyadMem takes 23.5 s per question to answer, close to AMA-Agent (20.7 s).
  • Construction is a one-time cost when multiple questions are asked on the same trajectory. Memory construction is question-independent, so the more questions per trajectory, the lower the per-question share. DyadMem's construction time is comparable to AMA-Agent's (228.5 s vs. 236.0 s per trajectory) and only about 1/6 to 1/5 of that of HippoRAG 2, STITCH, and ReadAgent-P.
  • Lower input tokens and latency than methods of similar accuracy. The baseline closest in accuracy is Long Context. Compared with it, DyadMem uses less total input per question on both benchmarks (44.6k vs. 142.9k; 37.3k vs. 54.1k); its non-cached input is about half on CAME (21.7k vs. 45.3k) and roughly equal on AMA (27.4k vs. 26.8k); and its latency drops from 73.9 s to 42.5 s per question. Methods with lower total input per question than DyadMem (AMA-Agent, Agentic Dense RAG, HippoRAG 2, MemGPT, and the RAG methods) are cheaper, but at least 31 points less accurate on CAME and at least 7.5 points less accurate on AMA.

Full cost statistics, including Qwen, will be added to the appendix of the revision.

Q4: Real interactive execution

The reviewer rightly points out that gains in retrospective QA do not necessarily translate into gains in online execution. To test this, we ran all 35 tasks on the Shopping Admin site of WebArena-Lite [1]. Here, memory is no longer built from pre-recorded trajectories but is updated incrementally as the agent executes, and the agent decides each next action based on its current memory. All methods use Gemma as the base model, with the same actor prompt, browser tools, and a 40-step action limit; only the history-memory module is replaced. Semantic judging uses Gemma. This setting places real demands on memory: 29 of DyadMem's 35 trajectories exceed 32k tokens, with a median trajectory length of about 116k tokens.

MethodSuccess rateAvg. steps
DyadMem17/35 (48.6%)23.8
AMA-Agent14/35 (40.0%)25.0
Agentic Dense RAG14/35 (40.0%)21.6
STITCH13/35 (37.1%)25.3
BM25 RAG12/35 (34.3%)26.5
ReadAgent-P12/35 (34.3%)24.1
Dense RAG11/35 (31.4%)25.3
HippoRAG 211/35 (31.4%)26.2
Long Context (256k)10/35 (28.6%)21.5
MemGPT10/35 (28.6%)22.5

In this small-scale preliminary experiment, DyadMem achieves the highest success rate among the 10 settings, providing initial evidence that it can also serve as the history memory of an online agent. Given the limited number of tasks, differences of a few tasks between methods require larger-scale evaluation to confirm, which we leave to future work. We will add this experiment to the appendix.

We thank the reviewer again. The controls, significance tests, cost statistics, and online experiments added following these suggestions have made the paper's conclusions more solid and helped us delineate the role of each component more precisely. We would be glad to address any remaining questions during the discussion period, and if these additions resolve your main concerns, we would kindly ask you to consider revisiting your score.

References

[1] Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Iong, Xixuan Song, Yifan Xu, Shudan Zhang, Hanyu Lai, Jiadai Sun, Xinyue Yang, et al. VisualAgentBench: Towards Large Multimodal Models as Visual Foundation Agents. In International Conference on Learning Representations (ICLR), 2025.

I appreciate the authors' efforts and find the additional evidence convincing. I have therefore decided to raise my score.

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

This paper investigates the problem of memory management within long-horizon agent trajectories. Specifically, it addresses how to maintain current task states, preserve historical event clues, and recover specific evidence lost during compression, all within a limited context window. The authors propose DyadMem, a framework that separately maintains episodic memory and a chronologically ordered event list. It controls context size through hierarchical re-compression while preserving mappings from events to their complete raw records. When answering questions, the system first generates a draft based on its memory, then searches or retrieves the raw records on demand. The paper compares this approach against eight other methods using two foundation models on the CAME-Bench and AMA-Bench benchmarks. Furthermore, the design is validated through component ablation studies and case analyses. The results support the overall effectiveness of the framework in long-trajectory question answering, particularly highlighting the complementary roles of the event list and the raw text re-reading mechanism.

Critical strengths

The paper effectively captures two distinct needs in long-trajectory memory: "maintaining global state" and "recovering local evidence." It organizes dual-zone memory, re-compression with preserved source mappings, and on-demand raw text re-reading into a cohesive workflow. Its innovative value lies primarily in the synergistic design of these mechanisms tailored for agent trajectories, as well as in the concrete implementation that maintains historical accessibility during continuous compression. The experiments are also quite compelling: DyadMem outperforms the compared methods in overall metrics across two benchmarks and two foundation models. On CAME-Bench with Gemma, removing episodic memory, the event list, or raw text re-reading reduces the F1 score by 9.01, 39.18, and 9.24 percentage points, respectively, providing direct evidence for the contribution of each component. Together, these results support the practical value of this work.

Critical weaknesses
  1. The independent contributions of the dual-zone division have not been sufficiently distinguished. Although the paper compares related frameworks such as ReadAgent-P, STITCH, and AMA-Agent, and uses ablation studies to demonstrate that both memory zones contribute to the current system, the analysis has limitations. Removing the event list simultaneously eliminates a large amount of historical content and source clues. According to Appendix A.3, event summaries and their mappings also participate in retrieval grouping and scoring; thus, this modification may further affect retrieval behavior. Consequently, the significant drop shown in Table 3 cannot be entirely attributed to the dual-zone division itself. Furthermore, since the situational memory targets only about 80 words, its poor standalone performance is also influenced by its limited information capacity and functional positioning. While existing results support the effectiveness of the complete framework, the unique benefits of the dual-zone organization compared to existing structured memories remain insufficiently clarified.
  2. There is a lack of sufficient evaluation on how the method consistently locates required evidence after source mappings are preserved. Section 3.3 explicitly states that structural checks do not guarantee the semantic integrity of summaries. However, clues needed to locate evidence may become blurred after repeated merging of summaries, and the retrieval mechanism itself involves candidate truncation. Appendices A.2–A.3 specify that each query selects at most two raw records from each event group, and note that pagination may return empty results before the reported total is exhausted. Therefore, when relevant evidence resides in other records within the same merged group, simply continuing pagination may not suffice to access it; the system would need to rewrite the query or select alternative IDs. The failure to find the price during the first read of representative records in Appendix E.2 also illustrates the distinction between source accessibility and successful localization. While existing ablations and cases provide positive evidence, they are insufficient to determine the reliability of this process under varying merge scales and re-compression counts.
  3. Some terminology in the main text and appendix does not clearly correspond. The main text uses "Situation Memory" and "Event List," whereas the prompts in Appendix A.4 use "Current Situation" and "Event Queue." It is recommended to clarify this correspondence upon first mention to allow readers to accurately map the method description to the actual prompts. This issue primarily affects clarity of presentation.
  4. The relationship between raw steps, storage blocks, and merged event groups needs further clarification. Sections 3.1–3.2 describe records and their IDs in terms of individual action-observation steps, while Appendix A.1 explains that the implementation may pack multiple dialogue turns into a single storage block, which are subsequently merged into event groups. The current explanation requires readers to infer the correspondence between objects of different granularities, which is particularly confusing regarding event IDs in the tool, representative IDs of merged summaries, and the scope of raw records they cover. Since source mapping is a crucial component of the method, these correspondences should be unified in the methodology section.
Questions for authors
  1. Clarification on dual-zone contributions: In the ablation that removes the Event List, beyond the historical summaries visible to the model, what specific changes occur in retrieval grouping, summary-involved scoring, and source localization? Is the budget freed by removing the module reallocated? These details would help determine how much of the drop in Table 3 reflects the dual-zone division itself, and how much reflects joint changes in information capacity and the retrieval mechanism. If there are results that control for these factors, please report them.
  2. Clarification on the evidence recovery mechanism: For a merged event group that covers multiple raw records, when the target evidence is not among the two records selected for the current query, does the tool provide a mechanism to continue traversing the remaining records within the group? If the system mainly relies on query rewriting or selecting alternative IDs, can the existing run logs report the recovery success rate in such cases, and its relationship to merged group size or the number of re-compression steps? This will affect the assessment of the practical reliability of the source-preservation mechanism.
  3. Clarification on record granularity and ID semantics: Please clarify the mapping among action–observation steps, packed storage blocks, and re-compressed event groups. What level of object do the IDs received by find_raw_events specifically refer to? When the model issues a read using the representative ID of a merged summary, does the tool return the single record corresponding to that ID, the entire storage block, or multiple records within the merged scope? It would be helpful to illustrate this with the example of reading e000307 in Appendix E.2, and to explain how this access behavior corresponds to the ID coverage relation in Equation (3).
4 comments
AuthorsOpenReview ↗

Response to Reviewer 2 (Part 1/3)

We thank the reviewer for the positive assessment of our method and experiments, and for the very careful reading of implementation details. The questions on pagination behavior and ID granularity helped us identify a genuine limitation and several unclear descriptions. We respond point by point below.

W1 / Q1: Confounds in the w/o Event ablation

The reviewer's analysis is accurate. Besides the historical event summaries becoming invisible, removing the Event List changes the following:

  • Retrieval grouping: every raw record becomes a singleton retrieval group;
  • Summary scoring: summaries no longer participate in retrieval scoring;
  • Source localization: no IDs appear in memory; the model can only read IDs returned by search results;
  • Budget: the Situation remains at about 80 words; the freed budget is not reallocated.

Therefore, −39.18 measures the overall effect of removing the entire Event List zone, and cannot be attributed entirely to the dual-zone division. To further examine the contribution of the dual-zone organization itself, we added a simpler summary-and-index control: a rolling summary plus a chronological index (Simple; see the General Response for the setup). It shares the raw backend, retrieval tool, draft-then-verify controller, and budget limits with DyadMem.

CAME/GemmaOverall F1SmallMediumLarge
Simple60.5381.3853.3031.20
DyadMem76.1785.5078.3848.08

DyadMem outperforms Simple by 15.64 F1, and by 25.08 and 16.88 on Medium and Large trajectories, respectively. Simple's summary is not subject to the ~80-word limit of the Situation and can record historical content as well. This suggests that a single rolling summary struggles to retain historical details across repeated updates, whereas the dual-zone design stores history separately in the Event List.

Existing ablations complement this control from two angles. w/o Situation keeps the per-record Event List, IDs, and the full retrieval machinery unchanged and removes only the state zone, yet F1 still drops by 9.01, indicating that the state zone contributes independently at the same history granularity and retrieval settings; together with Simple, this shows that both zones contribute. BM25 Retrieval keeps memory unchanged and only replaces query retrieval with per-record BM25 without summaries; F1 drops by only 0.93, indicating that retrieval grouping and summary scoring have a limited effect on overall results. In the revision, we will describe the above changes of w/o Event in Appendix D and add the Simple control.

AuthorsOpenReview ↗

Response to Reviewer 2 (Part 2/3)

W2 / Q2: Evidence localization within merged groups

For a merged group covering multiple raw records, find_raw_events offers two ways to access the remaining records in the group. The first is reading by ID: a merged Event List entry is annotated with the ID range it covers, and the agent can directly read any record in that range, up to 12 IDs per call. The second is query rewriting: when the query changes, the local ranking of records within the group changes as well, so records previously ranked lower may enter the top two. In Appendix E.2, the first read of the range's first record did not contain the price; after switching to an entity-name query, the agent obtained the naming record e000324 within the group and the price records e000393/e000395 outside it.

We thank the reviewer for catching this, and the reviewer's understanding of the pagination behavior is entirely correct: pagination alone cannot traverse the records within a group. Query retrieval first ranks event groups and selects at most the two highest-scoring records from each group, and only then applies max_results/match_offset pagination; therefore, continuing to page with the same query never returns records ranked third or lower within a group. total_matches counts the records covered by all positively scored groups, which also differs from the number retrievable through pagination. We originally capped each group at two records so that a "full-name record + later reference record" pair could be returned together, while preventing a single long event group from consuming a call's entire return quota; however, as the reviewer points out, this also means that lower-ranked target records in large merged groups cannot be reached through pagination.

The reviewer raises two concerns: whether localization cues become blurred after repeated merging, and whether candidate truncation prevents target records from being returned. We examined each with the fact-level audit.

Preservation of localization cues. Re-compression mainly omits specific values, while localization cues such as entity names and topics are almost always preserved: among 2,154 (fact, compression) pairs, cues disappeared entirely in only 2 pairs, both involving the same fact, which was eventually recovered by a global query.

Practical impact of the per-group cap. We replayed all 216 queries of the main-table run offline, and all results were reproduced exactly. The cap can only affect groups with 3 or more records: among 199 comparable (query, target group) pairs, the target group had only 1–2 records in 137 cases; exclusions for the remaining groups are as follows:

Number of raw records in the target group3–8≥9
Target source excluded by the per-group cap18/4416/18

Exclusions are concentrated in large merged groups of a single Large trajectory, whose relevant sources mostly went through 3–4 re-compressions. Of the 108 facts requiring correction (insufficient memory and an incorrect draft), 13 encountered such exclusions, and none subsequently obtained sufficient evidence; for 5 of them, lifting the cap on the corresponding group alone would have exposed all necessary evidence within the same return quota. This shows that the issue the reviewer identified does affect recovery in large merged groups. In the revision, we will add these statistics to Appendix A.2, clarify the meaning of total_matches, and discuss this issue in the limitations section.

AuthorsOpenReview ↗

Response to Reviewer 2 (Part 3/3)

W3: Inconsistent terminology

We thank the reviewer for pointing this out. The correspondence is as follows: Situation Memory corresponds to Current Situation in the prompts, and Event List corresponds to Event Queue. "Queue" in the prompts follows the naming in our implementation and refers to an event sequence appended in chronological order; it does not imply FIFO eviction, since summaries exceeding the capacity are merged through hierarchical re-compression. The prompts in Appendix A.4 are the exact text used in our experiments. To stay strictly consistent with the experiments, we prefer to keep them verbatim in the appendix, while adding a footnote at the first mention of these terms in the main text and a note before the prompts in Appendix A.4 explaining the correspondence.

W4 / Q3: Record granularity and ID semantics

We thank the reviewer for raising this point. DyadMem uses the record as its basic unit: the Raw Store stores records, IDs are record indices, and both the Event List summaries and the coverage relation in Eq. (3) are defined over records. A record corresponds to one interaction step, instantiated in each benchmark as follows:

  • AMA-Bench: one action–observation step is one record, consistent with the formalization in Sec. 3.1.
  • CAME-Bench: dialogues contain no explicit actions, so consecutive dialogue turns under the same timestamp are treated as one step and packed into one record (up to 8 turns and 2,048 tokens; Appendix A.1). For example, Step 307 (e000307) in Appendix E.2 consists of one user message and two agent replies: the user message plays the role of the observation and the agent replies that of the action.

Packing is thus only the way dialogues are mapped to interaction steps; the memory mechanism itself is unchanged. On this basis, the three levels of objects relate as follows:

  • Record: the basic unit in the Raw Store, each with a unique ID, e.g., e000307.
  • Initial event summary: one initial summary per record, retaining that record's ID.
  • Merged summary: re-compression merges adjacent summaries, and the merged entry is annotated with an ID range, e.g., [e000307..e000338]. In Eq. (3), ids(·) of this entry is the set of all record IDs in the range {e000307, …, e000338}, not just the first ID. The "representative ID" mentioned in the Case Analysis and Appendix E.2 of the original manuscript refers to the first ID of the range, e000307. It still points to record 307 itself and is not a new group identifier. Since this wording could suggest an ID representing the whole merged group, we have changed it to "the first ID of a merged range" in the revision.

Both ids(·) in Eq. (3) and the event_ids argument of find_raw_events refer to record IDs. When reading by ID, the tool returns only the requested record; an accompanying query only determines which part of a long record is returned and does not expand the entire merged range. The merged range serves two purposes: it is the grouping basis for query retrieval, and it tells the agent which IDs in the range can be read.

In the E.2 example, the agent first read e000307, the first record of the range, and the tool returned only record 307, a transitional exchange about the day-9 plan that contains no price. The subsequent name query returned the naming record e000324 and the price records e000393/e000395, yielding 108.65.

We agree with the reviewer that source mapping is a core component of the method and that these correspondences should be stated uniformly in the method section. In the revision, Sec. 3.2 will define records and their instantiation in each benchmark, specify the three levels of objects above, the unit of Eq. (3), and the semantics of ID reads, and include a granularity diagram using e000307 as an example; the packing rules in Appendix A.1 will refer back to this definition.

We thank the reviewer again for the support and detailed comments, which made the description of our mechanism much clearer and helped us identify the per-group retrieval cap as an area for improvement. We would be glad to address any remaining questions.

Thank you for the detailed response and additional analyses. The new summary-and-index control provides useful evidence for the dual-zone organization. The clarifications on record IDs and pagination, together with the retrieval audit, address my questions about the mechanism and quantify the remaining limitation of the per-group retrieval cap. These additions strengthen my positive assessment, and I maintain my rating of 3.

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

DyadMem combines a cumulative Situation summary with an ordered Event List, while retaining an immutable raw archive. Compression preserves identifiers and structural coverage, and a draft-and-verify procedure retrieves raw evidence when needed. The central idea is to use compressed memory as an organized access layer, rather than require the compressed representation to retain every answerable detail.

Critical strengths
  • Separating persistent situation information from chronological events is sensible for trajectory QA. Removing either zone hurts performance, and the large drop without the Event List supports the need for more than a single evolving summary.

  • The raw archive makes compression reversible. The paper also gives a useful example where information lost during repeated compression is recovered through raw search, and explicitly distinguishes identifier preservation from semantic completeness.

Critical weaknesses
  • The evaluation does not yet cleanly isolate the proposed memory organization from the draft-and-verify retrieval procedure. Removing an entire zone changes several capabilities at once. FIFO is relatively close overall, although the larger-model result shows a meaningful benefit from recursive compression; removing IDs also has mixed effects across models. These results make it important to test whether a simpler summary-and-index memory, given the same raw-search backend and verification controller, can obtain most of the gain.

  • Structural coverage is not the same as reliable recovery of forgotten information. The paper acknowledges this, but the central compression-and-recovery mechanism is demonstrated more clearly through examples than through a systematic test. Because raw search can bypass the identifier route, successful retrieval does not necessarily validate provenance-preserving compression. Another potential failure is a plausible but outdated draft that never triggers sufficient verification. The relevant question is whether the system recognizes and repairs information loss, not merely whether the raw record still exists.

Questions for authors
  1. With the raw archive, retrieval budget, and draft-and-verify controller held fixed, how much does the dual-zone organization improve over a simpler summary plus chronological index?

  2. Can you track facts across successive compression passes and measure recovery when their semantic cues disappear or become outdated? Please distinguish identifier-based recovery from global raw search, and report cases where verification fails to recognize that the initial draft needs correction.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1 (Part 1/2)

We thank the reviewer for recognizing the value of separating the Situation from chronological events and of the reversible compression enabled by the raw archive. We also appreciate the reviewer's precise distinction between identifier preservation and semantic completeness: this is exactly what we intended to convey in Sec. 3.3, but had not validated systematically. The reviewer's two questions target the parts of the method that most need validation, and we have added one experiment for each, following the reviewer's suggestions.

W1 / Q1: Benefit of the dual-zone organization with the raw backend, retrieval budget, and controller fixed

As suggested, we fixed the raw backend, retrieval budget, and draft-then-verify controller, and replaced the dual-zone memory with the simpler summary-plus-chronological-index alternative the reviewer described (Simple; see the General Response for the setup).

CAME/GemmaOverall F1SmallMediumLarge
Simple60.5381.3853.3031.20
DyadMem76.1785.5078.3848.08

With the same raw backend, budget, and controller, Simple is 15.64 F1 below DyadMem, with the gap concentrated on Medium (25.08) and Large (16.88) trajectories. This suggests that a rolling summary with a chronological table of contents cannot substitute for DyadMem's memory organization.

This control complements the existing w/o Situation ablation. w/o Situation keeps the per-record Event List, IDs, and retrieval settings unchanged and removes only the state zone, yet F1 drops by 9.01, showing that the state zone contributes independently at the same history granularity. Simple, in turn, shows that a single rolling summary serving as both state and history cannot replace a per-record history zone. Taken together, both zones contribute.

Regarding hierarchical re-compression and visible IDs, the reviewer notes that FIFO is relatively close to the full method overall and that removing IDs has mixed effects across scales. These two ablations measure different things and are best read separately:

  • Hierarchical re-compression (vs. FIFO). Its role is mainly structural: under a limited budget, old events are merged rather than discarded, so the Event List always covers the entire history. This difference only emerges when trajectories are long enough for the Event List to repeatedly reach its capacity. As the reviewer observed, recursive compression brings a clear benefit on the Large subset (FIFO 41.35 vs. full method 48.08).
  • Visible IDs. w/o Event IDs removes only the IDs rendered to the model; the internal source mapping is fully preserved, and retrieval grouping, the coverage relation, and reading raw records by ID all work as usual. This ablation therefore measures the benefit of the model seeing IDs directly, not whether source mapping itself is necessary. Its mixed effects across scales indicate that visible IDs by themselves play a limited role, while the much larger drop from removing the Event List entirely (−39.18) shows that the Event List contributes far more than displaying IDs.

These two are auxiliary designs of the Event List. The main design of DyadMem is the dual-zone memory, whose contribution is supported by the two complementary controls above, Simple and w/o Situation; removing the Event List or Raw Recall also reduces F1 by 39.18 and 9.24, respectively.

AuthorsOpenReview ↗

Response to Reviewer 1 (Part 2/2)

W2 / Q2: A systematic test of loss → recognition → localization → repair

The reviewer points out that the key question is whether the system recognizes and repairs information loss, not merely whether the raw record still exists. We fully agree, and conducted a fact-level provenance audit of the main-table run (CAME/Gemma, 373 questions) accordingly (automatic LLM annotation). The audit decomposes the reference answers into 972 facts with explicit raw-record support, and annotates each fact in the draft, in every snippet actually returned to the model by each tool call, and in the final answer. For 341 of these facts (price/rating facts and "current/latest value" facts), we additionally tracked their memory status before and after every compression. Answer correctness is judged by semantic equivalence.

Overall, 82.4% (801/972) of the facts are still directly supported by the compressed memory at answer time. For the remaining, insufficiently supported facts, verification raises the number of correct facts from 63 to 83. Across all facts, the number of correct facts increases from 698 in the draft to 760 in the final answer, with only 3 originally correct facts changed to incorrect (in the draft, 698 facts are correct, 273 incorrect, and 1 undeterminable, which is correct in the final answer). Results for each stage:

StageMetricResult
CompressionFact still supported by memory after one Work Space compression95.4% (1,485/1,557)
CompressionFact still supported by memory after one Event List re-compression72.7% (88/121)
OutcomeFacts lost during compression that are finally answered correctly70.3% (26/37)
RecognitionMemory insufficient and draft wrong → verification targeting this fact is triggered88.9% (96/108)
LocalizationSufficient evidence obtained after verification is triggered31.3% (30/96)
RepairAnswer corrected after sufficient evidence is obtained63.3% (19/30)
RiskOriginally correct draft facts changed to incorrect0.43% (3/698)

For the two compression rows, the denominator counts (fact, compression) pairs and includes only facts supported before the compression.

  • Recognition. Of the 108 facts with insufficient memory and an incorrect draft, 96 (88.9%) triggered verification targeting that fact. The remaining 12 (11.1%) are exactly the "plausible but wrong draft" cases the reviewer was concerned about. This failure mode does occur, and we will report it in the limitations section.
  • Localization and recovery path. Most of the loss occurs at localization: for 66 of the 96 facts that triggered verification, the retrieved snippets were still insufficient to support the answer. Among the 30 facts for which sufficient evidence was first obtained, 26 came from global queries, 2 from ID reads, and 2 from a combination of both. Raw recovery thus relies mainly on queries over the full Raw Store, with reading by ID serving as an auxiliary access path. Localization is the stage where the recovery pipeline loses the most; we will discuss this in the limitations section.
  • Net gain. Of all 273 facts that were wrong in the draft, 64 were corrected in the final answer, while only 3 of the 698 correct draft facts were changed to incorrect. This is consistent with the ablation: answering from memory alone without raw recall (w/o Raw Recall) reduces F1 by 9.24.
  • Outcome of lost facts. Of the 26 facts that were lost during compression but finally answered correctly, 12 obtained evidence through raw retrieval; for the rest, either a later memory update re-recorded the fact, or memory still retained a reference sufficient to determine the answer uniquely.
  • Cues that disappear or become outdated. Only one fact lost its semantic cues entirely, and it was recovered by a global query. This run contains no case where an outdated cue led to a wrong draft, so we cannot draw quantitative conclusions about this failure mode; we will discuss it in the limitations section.

This is consistent with the statement in Sec. 3.3: the structural invariant of re-compression guarantees that every raw record remains accessible, but not that summaries are semantically complete; omitted details remain in the corresponding raw records and can be recovered through retrieval. We will add this audit to the appendix.

We thank the reviewer again. These questions led us to test the compression-and-recovery mechanism systematically and to characterize the method's limitations more accurately. We would be glad to address any remaining questions during the discussion period, and if these additions resolve your main concerns, we would kindly ask you to consider revisiting your score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all three reviewers for their careful reading and constructive feedback. All reviewers found the motivation of separating global state maintenance from local evidence recovery sensible and the paper clear to follow, and R2 and R3 also found the empirical results and ablations convincing. The reviewers also raised several important questions: the contribution of the dual-zone organization itself, the reliability of evidence recovery after compression, the necessity of several smaller design choices, and computational cost and online execution. Following the reviewers' suggestions, we have added the following experiments:

  1. Rolling summary + chronological index control (Simple; R1, R2, R3): the simpler summary-and-index alternative suggested by the reviewers, sharing the Raw Store, retrieval tool, draft-then-verify procedure, and budget limits with DyadMem.
  2. Fact-level provenance audit (R1, R2): tracing 972 facts through each stage of loss during compression → recognition → localization → repair, plus an offline replay of all 216 queries to quantify the effect of the per-group cap.
  3. Trajectory-level paired bootstrap confidence intervals (R3) for all ablations in Table 3.
  4. Agentic Dense RAG, an iterative-retrieval baseline (R3).
  5. Token cost and controlled latency for all 10 methods (R3).
  6. Online execution on WebArena-Lite (R3).

Setup of the Simple control (not repeated in the individual responses). Simple processes raw records chronologically in segments, each consisting of complete records and up to 8k tokens. After each segment, the model rewrites a cumulative summary and writes a one-sentence gist of that segment; the gist is tagged with the segment's raw-ID range and appended in chronological order to form an index. In other words, Simple uses a single summary for both state and history, plus a coarse-grained table of contents, whereas DyadMem keeps the current state (Situation) and per-record event summaries (Event List) in separate zones. Both share the Raw Store, the find_raw_events tool, the draft-then-verify procedure (up to 3 recall rounds), and the budget limits (32k call context, 18k visible memory).

CAME/GemmaOverall F1SmallMediumLarge
Simple60.5381.3853.3031.20
DyadMem76.1785.5078.3848.08

The reviewers' comments also helped us identify places where the paper was underspecified or overstated. In the revision, we will: (i) add confidence intervals to Table 3, clarify what the hierarchical re-compression and visible-ID ablations each measure, and reduce the emphasis on named-anchor expansion and the retrieval backend in the main text; (ii) discuss the evidence-localization stage in the limitations section; (iii) define record granularity and ID semantics in a unified way in Sec. 3.2; and (iv) add all of the above experiments to the appendix.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.