acceptodds
Under review as a conference paper at ICLR 2027

CoMATE: Consolidating Memory Across Trajectories for Long-Horizon Interactive Exploration on ARC-AGI-3

Abstract

As agents take on increasingly complex and demanding tasks, reliable long-horizon performance is becoming essential but remains difficult to achieve. This difficulty is evident in long-horizon interactive exploration, where ARC-AGI-3 remains very hard for frontier models in official direct-play evaluations. Parallel exploration at test time can improve performance by generating complementary experience, but shared notes alone do not establish which claims are supported, when they apply, or how conflicting accounts should be reconciled. We introduce CoMATE, an evidence-grounded memory consolidation framework. A fresh-context model examines agents' written memories and observer-view trajectories, producing collective entries with supporting evidence, epistemic status, and applicability conditions. Agents query this knowledge during subsequent exploration. Because results on ARC-AGI-3 depend heavily on the agent harness, we compare memory treatments under a shared parallel search protocol. Level-wise branching, checkpoint handoff, and retries are common to all conditions, so the no-consolidation baseline also retains a full parallel group at each new level. Under this protocol, CoMATE raises RHAE from 55.1 to 80.1 with six GPT-5.5 agents on the 25 public games, completing 22 games and 171 of 183 levels; nine agents reach 80.9 RHAE and 173 levels. With GPT-5.6 Sol, it raises RHAE from 87.6 to 97.0 and completes all 25 games and 183 levels. Action- and token-budget curves account for both gameplay and memory-related inference, while automated and human audits separately assess memory correctness and citation support. These results support cross-trajectory consolidation on ARC-AGI-3 and its potential for broader long-horizon interactive exploration.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Multi-Agents, Long-Horizon Exploration

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KyUWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. the paper introduces CoMATE, a harness or multiagent harness for agents interacting and consolidating previous interaction histories into reusable shared knowledge pools accessible by all agents for future game exploration. The stored knowledge or interaction histories mainly record discovered world/game rules and when they can be applied.
  2. the core contribution part is the consolidation of previous histories while other parts of the harness are mostly common and presented as baselines. The consolidation is done by a separate agent verifying and consolidating the agents that operate in the game's notes into shared findings with grounded evidence and also some uncertain discoveries.
  3. the paper mainly tries to solve ARC-AGI-3 benchmark and shows improvements over baseline harness without knowledge consolidation. The baseline harness was adapted from a prior released harness.
Critical strengths

The paper demonstrated that with memory consolidation across agents and maintaining a shared pool of grounded knowledge of the game world, it is beneficial to the multiagent system for future exploration and operation in the game world. This is the core contribution in my opinion, and the paper mainly compares CoMATE versus B6 (an adapted harness from an existing released harness). The results are supporting that shared memory consolidation is useful for multiagent systems and thus is what makes me think the paper has some contributions overall.

Critical weaknesses
  1. Only testing on ARC-AGI-3 makes the claim that sharing and consolidating memories across agents (with specific checks from an external agent instead of directly trusting the single agent in the game worlds) in exploring an open world less convincing on its generalizability across diverse tasks. Even though the results on ARC-AGI-3 is supporting the main claim, limiting in 1 type of environment makes the harness design less convincing.
  2. The paper already cites exiting works such as G-Mem [1] and Symbolica [2] and the claimed contribution is on whether the shared knowledge should be consolidated by another agent instead of trusting the agent that acts in the games. The claim is a good direction to study but the main experiments are not substantial in showing this. The incremental improvements may need better grounding in the existing literatures. Also, related work [3] has similar knowledge consolidating mechanism though is more grounded in the environment and thus making the grounding more verifiable, so the contribution claim may need to be reconsidered.

[1] Zhang, Guibin, et al. "G-memory: Tracing hierarchical memory for multi-agent systems." Advances in Neural Information Processing Systems 38 (2026): 12988-13018. [2] Symbolica. From 0% to 36% on day 1 of ARC-AGI-3. Symbolica Blog, March 2026a. URL https://www.symbolica.ai/blog/arc-agi-3. [3] Xu, H., Hu, J., Zhang, K., Yu, L., Tang, Y., Song, X., ... & Shi, B. (2025). Sedm: Scalable self-evolving distributed memory for agents. arXiv preprint arXiv:2509.09498.

Questions for authors
  1. Can you demonstrate the CoMATE also works well on some other kind of environments? I understand the focus is on ARC-AGI-3 here but to demonstrate the usefulness of memory consolidation part do need diverse settings in my opinion.
  2. Can you discuss more on the contribution against prior works such as [1,2,3] in knowledge consolidation? More specifically, how would the authors ground the work against SEDM for evidence grounding in traces or environments?

[1] Zhang, Guibin, et al. "G-memory: Tracing hierarchical memory for multi-agent systems." Advances in Neural Information Processing Systems 38 (2026): 12988-13018. [2] Symbolica. From 0% to 36% on day 1 of ARC-AGI-3. Symbolica Blog, March 2026a. URL https://www.symbolica.ai/blog/arc-agi-3. [3] Xu, H., Hu, J., Zhang, K., Yu, L., Tang, Y., Song, X., ... & Shi, B. (2025). Sedm: Scalable self-evolving distributed memory for agents. arXiv preprint arXiv:2509.09498.

3 comments
AuthorsOpenReview ↗

Response to Reviewer KyUW

Thank you for recognizing that "maintaining a shared pool of grounded knowledge of the game world" benefits the multiagent system, and for raising the question of whether shared knowledge "should be consolidated by another agent instead of trusting the agent that acts in the games."


1. Other environments. Following your suggestion, we added an environment and tested CoMATE on Jericho text adventure games, which differ from ARC-AGI-3 in both observation modality and task structure. In these human-written games (Hausknecht et al., 2020), the agent reads text, types commands, and scores by uncovering mechanisms hidden behind long chains of exploration. We keep our protocol: stages end at the next score gain, six agents start each stage from the winner's state, consolidation follows every round, and a stage all agents fail gets one more attempt and, if that also fails, is replayed from another agent's endpoint of the previous stage.

Jericho, 56 games, GPT-5.5Share of maximum scoreStages completed
Baseline44.7%438
CoMATE54.3%568

CoMATE passes the stage at which the baseline stops in 46 of the 56 games. On the 18 hardest games, whose reference walkthroughs exceed 300 commands, it raises the share from 16.4% to 27.0%.


2. Do the experiments support the claim? We agree this is the central question. Table 3 isolates it, with the backbone, agents, harness, budgets, and update points fixed:

  • Trusting the acting agent (Individual memory, with labels such as CONFIRMED assigned by each writer, shared with all agents): 51.0 RHAE, below no sharing (55.1).
  • A separate consolidator that does not check (Unstructured entries; same inputs, storage, and retriever as CoMATE): 72.8.
  • A separate consolidator that checks each claim against the trajectories (CoMATE): 80.1; every one of five CoMATE runs exceeds every unstructured-entry run.

Two new results strengthen this. First, claims their writers labeled CONFIRMED are factually supported only 76.3% of the time, versus 89.9% for CoMATE's entries. Second, published memory methods fall short of CoMATE under the same protocol, with each replacing only CoMATE's consolidation and retrieval and building memory as in its paper:

Cross-agent memory (GPT-5.5)RHAE/LAELevelsClears
None (B6)55.1/67.5142/183 (92)13/25 (6)
ReasoningBank (Ouyang et al., ICLR 2026)63.0/73.1151/183 (100)16/25 (7)
G-Memory (Zhang et al., NeurIPS 2025)56.9/69.2145/183 (94)13/25 (6)
CoMATE80.1/85.9171/183 (132)22/25 (12)

On the 12 games B6 does not clear, CoMATE clears 9, versus 3 for ReasoningBank and 1 for G-Memory.


3. Relation to SEDM, G-Memory, and Symbolica.

  • SEDM (Xu et al., 2025) grounds a memory item in measured utility: it replays the item's stored task context with and without the item and admits the item when the reward gain, net of latency and token costs, exceeds a threshold; it is evaluated on LoCoMo, FEVER, and HotpotQA. CoMATE grounds a claim in the recorded trace: each entry cites the observations and feedback that support it, with its status and the conditions under which it holds. In ARC-AGI-3, reward arrives only at level completion and a mechanism often matters only at a later level (as with the energy blocks in LS20), so the utility of a single claim is hard to measure by replaying the current level; checking it against cited records makes every entry auditable instead (Appendix C). We see the two forms of grounding as complementary.
  • G-Memory organizes multi-agent experience in a three-tier hierarchy of insight, query, and interaction graphs, with insights summarized from trajectories and task outcomes; it is evaluated on ALFWorld, SciWorld, PDDL, HotpotQA, and FEVER. CoMATE consolidates the experience of parallel agents within one environment, checks each claim against action–observation records, and keeps conflicts explicit. Under our protocol, G-Memory reaches 57.0 RHAE, versus 80.1 for CoMATE (table above).
  • Symbolica's harness shares one memory store between an orchestrator and its subagents, with confirmed/hypothesis labels set by each writer and no check against the trajectory; our Individual memory condition tests this directly (51.0 versus 55.1 without sharing).

We hope these clarifications and results address your concerns; if so, we would be grateful if you would consider raising your score.

Dear Reviewer KyUW, thank you again for your review. As the discussion period ends on November 18, we wanted to check whether our response addresses your concerns. We would be happy to clarify anything further. We would also be grateful if you could consider raising your score based on our response.

The response mostly addressed my concerns, and I have raised my score.

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

The paper proposes CoMATE, a memory consolidation framework cross-checking multiple agents' individual memories to produce collective memory entries with supporting evidence, epistemic status, and applicability conditions. The authors present action- and token-budget curves for memory-related inference, along with automated and human audits of memory factual correctness and citation (evidence) entailment. Evaluations reported on the 25 public ARC-AGI-3 games, comparing memory treatments under a shared parallel search protocol (level-wise branching, checkpoint handoff, and retries)

Critical strengths
  • Extending the idea of agents' notes jointly visible, to carry supporting evidence and applicability conditions, targeted for the problem of conflicting interpretations and accumulated errors in long-horizon exploration.
  • Performance gains are evaluated against both environment-action budgets and token budgets
  • The paper reports gains by controlled ablation for memory sharing, summarization, and checking claims against recorded trajectories (Table 3).
Critical weaknesses
  • The reported evaluation focuses on the public ARC-AGI-3 games and a specific agent/harness configuration, so generalization to other environments or real-world long-horizon tasks requires further validation.
  • For strong backbones (e.g., GPT-5.6 Sol), the remaining headroom is limited, so a broader analysis of how the effect varies across models and tasks would be useful.
  • The process of checking the new entry against the collective memory solely depends on LLM’s qualitative reflection. No verification by the direct signal (e.g. reward) from the environment.
Questions for authors
  • Answers for generalizability concerns in the weakness?
  • Does CoMATE yield comparable gains in environments where the game structure or observation modality differs?
  • When consolidation adopts an incorrect claim or omits correct information, how does this affect the behavior of agents in subsequent exploration?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Teex

Thank you for highlighting that CoMATE targets "conflicting interpretations and accumulated errors in long-horizon exploration," that its gains are evaluated against both action and token budgets, and that Table 3 provides a controlled ablation.


W1, Q1, Q2. Generalization. Although our claims concern ARC-AGI-3, where each of the 25 games is a new environment, we have now also tested CoMATE on a different kind of environment to answer your questions.

  • Within ARC-AGI-3, CoMATE was developed blind to individual games, with one method, prompt set, and protocol. With GPT-5.5 it improves 19 of the 25 games and ties the other 6 at 100 RHAE; it improves all three backbones.
  • A new environment. Jericho (Hausknecht et al., 2020) consists of human-written text adventure games, which differ in both observation modality and task structure: the agent reads text, types commands, and scores by uncovering mechanisms hidden behind long chains of exploration. We keep our protocol: stages end at the next score gain, six agents start each stage from the winner's state, consolidation follows every round, and a stage all agents fail gets one more attempt and, if that also fails, is replayed from another agent's endpoint of the previous stage.
Jericho, 56 games, GPT-5.5Share of maximum scoreStages completed
Baseline44.7%438
CoMATE54.3%568

CoMATE passes the stage at which the baseline stops in 46 of the 56 games. As on ARC-AGI-3, CoMATE also improves the hardest games: on the 18 games whose reference walkthroughs exceed 300 commands, it raises the share from 16.4% to 27.0%. These results suggest that the benefit of evidence-grounded consolidation is not specific to ARC-AGI-3's visual games.


W2. Headroom with strong backbones. Gains near the ceiling are the hardest to obtain, and the share of the remaining gap that CoMATE closes grows with backbone strength: 10.8% for GPT-5.4 (25.2 → 33.3), 55.7% for GPT-5.5 (55.1 → 80.1), and 75.8% for GPT-5.6 Sol (87.6 → 97.0). With Sol, CoMATE completes the 6 levels and 2 games that B6 leaves unfinished, and raises full-score clears from 14 to 23. Across games (Table 2, GPT-5.5), CoMATE mainly finds more efficient solutions on the 13 games B6 already clears (79.4 → 95.8 RHAE) and mainly adds coverage on the other 12 (28.7 → 63.1 RHAE; 0 → 9 clears).


W3. Checks rely on the LLM rather than an environment signal. You are right that the check is made by an LLM rather than by a reward signal; what distinguishes it from qualitative reflection is what the claim is checked against.

  • The consolidator checks each claim against the observer-view trajectory, that is, the observations before and after each action and the resulting feedback, and attaches the records that support it, rather than relying on agreement among the agents' accounts (Section 2.2).
  • ARC-AGI-3's only reward, level completion, triggers consolidation and gives the best-performing agent provisional priority on unresolved findings, without overriding contradictory evidence.

Level completion alone cannot verify a claim: a level can be passed with a partly wrong interpretation, and an agent that fails may have observed a correct mechanism (the LS20 example in the Introduction). The human audit confirms the checks: of 500 entries, 95.6% were judged factually supported and 90.2% had citations that entail the finding (Appendix C).


Q3. Effect of incorrect or missing claims. Three mechanisms keep an incorrect or missing entry from derailing exploration:

  1. Entries inform rather than dictate. They are returned as findings with evidence and status, not action plans, and agents judge them against the current observation (Section 2.3).
  2. Later observations revise entries. During exploration, 20.4% of entries were revised in their evidence, scope, or status (Appendix C).
  3. Stalled exploration recovers missing knowledge. On the 7 levels the initial GPT-5.5 pass leaves blocked, consolidating the failed attempts resolves 5, versus 1 for a fresh retry and 0 for continuing (Table 5).

We hope these answers and results address your concerns; if so, we would be grateful if you would consider raising your score.

Dear Reviewer Teex, thank you again for your review. As the discussion period ends on November 18, we wanted to check whether our response addresses your concerns. We would be happy to clarify anything further. We would also be grateful if you could consider raising your score based on our response.

Dear Reviewer Teex, we would like to gently follow up as the discussion period is approaching its end. We would be very grateful if you could let us know whether our response has addressed your concerns. If so, we would sincerely appreciate your reconsideration of our score. We are also happy to provide any further clarification if needed. Thank you again for your time and consideration.

Reviewer_v13POpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CoMATE is an evidence-grounded memory consolidation framework. A fresh-context model examines agents’ written memories and observer-view trajectories, producing collective entries with supporting evidence, epistemic status, and applicability conditions. Agents query this knowledge during subsequent exploration. With GPT-5.6 Sol, it raises RHAE from 87.6 to 97.0 and completes all 25 games and 183 levels.

Critical strengths
  1. The paper proposes an approach to cross-trajectory memory consolidation. Checking agents’ written memories against trajectory evidence and recording epistemic status and applicability conditions provides a principled way to combine complementary findings and handle conflicting interpretations.

  2. The shared parallel search protocol and progressively enhanced memory conditions help isolate the contribution of memory consolidation. Consistent improvements across model backbones, together with automated and human memory audits, provide convincing evidence of the method’s effectiveness.

Critical weaknesses
  1. The comparison with B6 and the progressively enhanced ablations help isolate the contribution of memory consolidation. However, direct comparisons with representative existing memory methods are missing, leaving CoMATE’s relative performance insufficiently established.

  2. The paper reports total token usage, including memory-related calls, but does not report end-to-end runtime. It remains unclear whether synchronous consolidation and repeated retrieval introduce critical-path latency that offsets the time saved by reducing exploration steps.

  3. The memory audits assess factual correctness and citation support, but do not directly quantify how consolidation improves knowledge quality. A comparison of the same claims before and after consolidation is needed to distinguish complementary knowledge acquisition, error correction, and conflict resolution.

Questions for authors
  1. Could the authors include representative existing memory methods as baselines under the same parallel search protocol?

  2. Could the authors report the wall-clock time each method takes to complete games under the same concurrency configuration, with a breakdown into gameplay, trajectory analysis, memory consolidation, and retrieval? In particular, how does the critical-path latency introduced by synchronous consolidation and repeated retrieval compare with the time saved through fewer exploration steps?

  3. Could the authors compare the same set of claims before and after consolidation, separately reporting the numbers and proportions of newly incorporated, evidence-supported complementary findings, erroneous claims corrected or excluded, and conflicts resolved?

2 comments
AuthorsOpenReview ↗

Response to Reviewer v13P

Thank you for the constructive review and for noting that checking memories against trajectory evidence "provides a principled way to combine complementary findings and handle conflicting interpretations." We followed all three suggestions.


Q1. Existing memory methods under the same protocol. CoMATE outperforms both ReasoningBank (Ouyang et al., ICLR 2026) and G-Memory (Zhang et al., NeurIPS 2025). Each replaces only CoMATE's consolidation and retrieval and builds memory as in its paper, with level completion as the success signal; the backbone (GPT-5.5), harness, budgets, retries, and other settings are identical.

Cross-agent memory (GPT-5.5)RHAE/LAELevelsClears
None (B6)55.1/67.5142/183 (92)13/25 (6)
Individual memory (Symbolica's native memory)51.0/63.4145/183 (77)13/25 (6)
ReasoningBank63.0/73.1151/183 (100)16/25 (7)
G-Memory56.9/69.2145/183 (94)13/25 (6)
CoMATE80.1/85.9171/183 (132)22/25 (12)

Both published methods improve on no sharing, but CoMATE leads them by 17.1 and 23.2 RHAE. On the 12 games B6 does not clear (Hard-12, Table 4), CoMATE clears 9, versus 3 for ReasoningBank and 1 for G-Memory.


Q2. Wall-clock time and critical-path latency. On the 142 levels both methods complete, CoMATE takes less time per level than B6, memory processing included. Both run under the same concurrency; the table gives mean wall-clock minutes per level on the critical path:

Mean per level (min)GameplayTrajectory analysisConsolidationRetrievalTotal
B614.5–––14.5
CoMATE10.12.21.20.313.8

Fewer exploration steps save 4.4 minutes of gameplay per level, more than the 3.7 minutes that memory processing adds to the critical path.


Q3. The same claims before and after consolidation. Consolidation adds findings that only one agent held, filters errors, and keeps conflicts explicit. For every consolidation event, we split the agents' notes into atomic claims, judged each against the trajectories, and traced it into the collective memory:

CategoryCountShare
Complementary findings: supported entries that at most one agent had noted before2,11338.3% of entries
Contradicted claims: corrected / excluded931 / 1,16441.4 / 51.7% of contradicted claims
Cross-agent conflicts: resolved by evidence / scoped by conditions / kept unresolved407 / 158 / 7563.6 / 24.7 / 11.7%

Factual support rises from 72.4% for the agents' raw claims, and 76.3% for those their writers labeled CONFIRMED, to 89.9% for the consolidated entries. A human check of 100 classifications agrees with 91.0% of the automated ones.


We hope these additions address your questions; if so, we would appreciate your considering a higher score.

Thanks for the detailed reply! I maintain my score and still recommand acceptance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all three reviewers for their careful and constructive reviews, and we are encouraged that they found our approach "principled" (v13P), our evaluation well controlled (v13P, Teex), our resource accounting informative (Teex), and our results supportive of the central claim (KyUW). Following their suggestions, we ran new experiments and analyses:

  1. Published memory methods under our protocol (v13P, KyUW). With the same GPT-5.5 backbone, agents, budgets, and memory update points, CoMATE exceeds ReasoningBank and G-Memory by 17.1 and 23.2 RHAE and, on the 12 games B6 does not clear, clears 9 versus 3 and 1.
  2. The same claims before and after consolidation (v13P). Factual support rises from 72.4% for the agents' raw claims to 89.9% for the consolidated entries.
  3. Wall-clock time (v13P). On the levels both methods complete, CoMATE takes 13.8 minutes per level versus 14.5 for B6, memory processing included.
  4. A new environment (Teex, KyUW). On Jericho text adventure games, CoMATE raises the share of the maximum score reached from 44.7% to 54.3% over the same-protocol baseline.
  5. Clarifications (Teex) on how claims are checked against the environment's recorded observations and how errors are contained, and positioning (KyUW) against SEDM, G-Memory, and Symbolica's harness.

We respond to each reviewer below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.