acceptodds
Under review as a conference paper at ICLR 2027

MemCo: Memory-Centric Collaboration for Generalizing LLM Agents to Unseen Environments

Abstract

Large language model (LLM) agents interact with environments through sequential observation, action, and feedback. While memory enables agents to reuse past experience, existing approaches often construct and maintain memory for individual agents independently, making it costly to collect sufficient trajectories and limiting generalization to unseen environments. Sharing experience accumulated by multiple agents interacting with different environments can reduce repeated exploration. In this work, we ask: how can shared experience be organized into useful memory and effectively used to support another agent’s current decision? We propose MemCo, a memory-centric collaboration framework for sharing and reusing agent experience. MemCo maintains complementary local and global memory spaces, preserving environment-specific experience locally while reusable records abstracted from local trajectories are aggregated across agents and promoted to global memory based on accumulated evidence. During online interaction, MemCo retrieves and composes memories from both levels based on the agent’s current state and decision phase, grounding shared experience in the current environment to support the next decision. Extensive studies on ALFWorld, PDDL, FEVER, and ScienceWorld show that MemCo improves task success under evaluation across seen and unseen environments, as well as reduces redundant exploration compared with the state-of-the-art methods. Our code is available at project.

Primary area:
applications to robotics, autonomy, planning
Keywords:
LLM agents, agent memory, multi-agent collaboration, interactive decision-making, episodic 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_ReQPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MemCo studies how independently acting LLM agents can reuse experience collected in different environments. It retains concrete experience locally, abstracts transferable records into shared global memory, and uses a Wilson lower confidence bound to determine promotion eligibility. At decision time, records are retrieved, checked for compatibility, grounded in the current environment, and composed according to state, task phase, and goal. Experiments across four benchmarks evaluate task performance and interaction efficiency, with ablations examining memory scope, promotion, and retrieval.

Critical strengths

Strength1: The framework explicitly handles when shared experience is applicable. Abstraction preserves operational conditions, while decision-time composition binds abstract roles to current objects and available actions (§§3.4–3.5). This connects transferable procedures with concrete execution requirements. The memory-scope and fixed-concatenation comparisons support the value of selectively using shared experience rather than simply exposing the policy to more records (Fig. 4).

Strength2: The empirical benefits extend across different models and task interfaces. Tables 1–2 show improvements beyond a single backbone or environment, while also revealing settings where MemCo is not the strongest method. The interaction-budget comparison further supports its ability to complete tasks with fewer execution opportunities (Fig. 3), making the contribution relevant to agents with constrained interaction budgets.

Critical weaknesses

Weakness1: The aggregate evaluation leaves the source of transfer gains partly unresolved. The merged test set includes both an agent’s acquisition domain and other domains (§4), but the reported results do not separate these cases. G-Memory and the memory-scope ablations provide useful comparisons; however, a simple pooled-record control would more directly test what the proposed local–global organization adds beyond access to broader source experience.

Weakness2: The promotion ablation does not fully separate evidence-aware selection from pruning strength. Replacing Wilson promotion with the empirical support ratio at the same threshold changes both record membership and global-memory size (Figs. 3–4). Similar injected context sizes help exclude prompt length as the sole explanation, but an equally selective admission rule remains a relevant comparison.

Questions for authors

Question1: Could you separate results on each agent’s acquisition domain from those on other domains? A pooled-record control using the same source trajectories and retrieval/composition procedure would additionally help isolate the benefit of local–global organization.

Question2: Could you compare Wilson promotion with selecting the same number of eligible records by empirical support rate, keeping retrieval and composition unchanged? Repeated random tie-breaking where necessary would help distinguish evidence-size-aware selection from generic pruning.

4 comments
AuthorsOpenReview ↗

Response to the Reviewer (1/2)

The suggestion to compare promotion methods at a matched record count directly motivated a valuable controlled experiment. We are grateful for this guidance and the constructive questions that helped strengthen our empirical evaluation.

Q1. Could results be separated by each agent’s acquisition domain and other domains? Could a pooled-record control with the same source trajectories and retrieval/composition procedure clarify the benefit of local–global organization?

A1. Thank you for the great suggestions. Following your suggestion to separate the results based on domains, we have added the requested domain breakdown and pooled-record control on ALFWorld with Qwen3-32B to clarify where the transfer gains occur and what local–global organization contributes beyond access to broader experience.

Regarding your 2nd question on the benefit of local-global organization, the submitted ablations support both combining local and global memory and selectively using their contents. Full outperforms Local-only and Global-only in Table 10. With the same underlying local and global memories, Full improves accuracy from 86.69% to 97.40% on Familiar Layouts and from 81.79% to 97.86% on Held-out Layouts compared with Fixed L5+G5, respectively. This comparison supports query-guided retrieval and composition with stored experience held fixed.

The added comparison separates acquisition-domain and other-domain performance and tests Full against directly pooled local records. Own/Other is defined relative to each agent’s acquisition domain within each layout split. Pooled uses the same frozen source memories and preserves record provenance, evidence, and each agent’s local graph. It omits global abstraction, cross-agent evidence aggregation, and promotion, while keeping retrieval, applicability checks, grounding, composition, and budgets unchanged. The largest observed accuracy gains occur on other-domain tasks under Held-out Layouts. Full improves accuracy by 11.90 percentage points over Local-only and 4.29 points over Pooled, while reducing average steps from 15.88 and 12.45 to 11.17. These results support constructing transferable global memory beyond directly pooling local experience. Full does not improve own-domain accuracy over Local-only or Familiar accuracy over Pooled.

The table reports accuracy (%) / average steps, using the original frozen-snapshot Full reference. Each condition evaluates 147 tasks under four agents’ memories. Gains are Full minus Local-only / Pooled, respectively.

LayoutDomainEvaluations per settingLocal-only Acc. / StepsPooled Acc. / StepsFull Acc. / StepsAccuracy gains (pp)
FamiliarOwn7798.70 / 10.3998.70 / 10.3597.40 / 9.58−1.30 / −1.30
FamiliarOther23195.67 / 12.2997.84 / 10.3897.40 / 9.79+1.73 / −0.43
Held-outOwn7098.57 / 11.8495.71 / 11.7495.71 / 12.17−2.86 / 0.00
Held-outOther21086.67 / 15.8894.29 / 12.4598.57 / 11.17+11.90 / +4.29
AuthorsOpenReview ↗

Response to the Reviewer (2/2)

Q2. Could Wilson promotion be compared with selecting the same number of eligible records by empirical support rate, keeping retrieval and composition unchanged and repeating random tie-breaking where necessary?

A2. Thank you for suggesting this controlled comparison. Our original experiments show higher accuracy under Wilson promotion than empirical-support promotion, while retaining fewer records. At threshold 0.35, Wilson retains 426 rather than 736 records and achieves Familiar/Held-out accuracies of 97.40%/97.86%, compared with 95.13%/96.43% for empirical support (Table 10).

In the repeating equal-count experiment, both methods select 426 records from the same pool of 749 eligible records, obtaining results consistent with the above finding. We constructed this pool from the four agents’ frozen local memories by applying abstraction, merging, and checks for structural validity and informativeness. Wilson promotion retained 426 of these records at the original threshold of 0.35, so we also selected the top 426 by empirical support rate. Because 735 records tied for the highest empirical support rate (1.0), we randomly selected 426 from this tied group in each of three repeats. We used no secondary ranking by evidence count. We held the local memories, record evidence, retrieval, grounding, and composition fixed. At this default bank size, Wilson achieves higher accuracy on both layout splits than the mean of the three empirical selections. We use the original Wilson results from Table 10 as the reference. Empirical tie-breaking uses selection seeds 101–103, and all three resulting memory banks are evaluated with inference seed 42 on the same agent–task pairs. We report each empirical selection and the mean ± SD, where SD reflects variation across the three record selections rather than independent memory-acquisition runs.

Promotion criterion / referenceRecordsFamiliar Acc. (%)Familiar StepsHeld-out Acc. (%)Held-out Steps
Wilson42697.409.7497.8611.42
Empirical top-K repeat142696.1010.5196.7912.33
Empirical top-K repeat242696.7510.3994.6412.51
Empirical top-K repeat342696.1010.6297.1411.66
Empirical top-K: mean ± SD42696.32 ± 0.3710.51 ± 0.1296.19 ± 1.3512.17 ± 0.45

Thank you for the detailed rebuttal and the additional controlled experiments. The domain-wise breakdown and pooled-record control clarify the source of the transfer gains, while the matched-record-count comparison with repeated tie-breaking better isolates the effect of Wilson-based promotion. These results adequately address my concerns. I therefore maintain my positive rating of 3.

Thank you very much for your encouraging feedback and for recognizing our additional efforts in addressing your concerns. We sincerely appreciate your positive assessment and continued support for our work.

Reviewer_eUKrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a memory-centric collaboration framework MemCo that enables LLM agents to share and reuse experience across different environments. The framework maintains local memory for environment-specific knowledge and global memory for reusable patterns promoted through evidence-based selection, while dynamically retrieving and grounding relevant memories to guide decisions. Experiments on ALFWorld, PDDL, FEVER, and ScienceWorld demonstrate improved task success and generally reduced interaction steps, particularly in unseen environments.

Critical strengths
  • MemCo handles the problem that experience from another environment can contain both useful procedures and misleading instance-specific details.
  • The comparisons with local-only, global-only, empirical-support promotion, fixed concatenation, and alternative retrieval strategies are relevant.
Critical weaknesses
  • The evaluation does not sufficiently separate the proposed architecture from access to a broader experience pool. For example, ALFWorld uses four acquisition domains with 100 training tasks each. An individual local memory contains experience from one domain, while the global memory can draw on all four. Section 4 and Appendix F do not clearly establish whether baselines receive the same source trajectories, total acquisition budget, environment information, number of few-shot examples etc.

  • Important implementation details that appear central to performance are described semantically rather than operationally. These include inducing workflows and preconditions, normalizing observations, identifying task phases, mapping entities to transferable roles, and determining whether a record is informative enough to share.

Questions for authors

PLease address Weknesses above.

3 comments
AuthorsOpenReview ↗

Response to the Reviewer (1/2)

Thank you for carefully examining the evaluation protocol and implementation details. Your concern about acquisition fairness prompted us to examine more directly the contribution of memory design beyond access to a broader experience pool.

W1. Does the evaluation separate MemCo’s memory design from access to a broader experience pool, and are the acquisition conditions comparable across methods?

A1. Thank you for the great questions. Yes, our evaluation separates MemCo memory design from access to a broader experience pool and the acquisition conditions are comparable across methods.

All memory-based methods receive the same training-task set within each benchmark, with at most 30 environment steps per training task. MemCo constructs global memory from its agents’ local memories without additional training tasks. This applies to ALFWorld, PDDL, FEVER, and ScienceWorld. Each method independently executes the shared training-task set once per acquisition run and constructs memory from its own interaction trajectories following its original procedure. The common task set and per-task step limit give all methods the same maximum total environment-interaction budget. All evaluated methods use one static few-shot example and access the same environment interface. MemCo’s source-domain assignment organizes its agents’ local experience, while baselines retain access to the full shared training-task set.

With Qwen3-32B, MemCo achieves the highest accuracy or success rate across all four benchmarks among ICL, A-Mem, MemSkill, MemRL, and G-Memory (Tables 1–2). G-Memory provides hierarchical graph-based memory for multi-agent systems, so the comparison includes an existing shared-memory approach.

The ablation reported in the submitted version further shows gains from query-guided selection and composition while keeping the local and global memories fixed. On ALFWorld with Qwen3-32B, Full MemCo and Fixed L5+G5 use the same underlying memories, but Full improves Familiar and Held-out accuracy by 10.71 and 16.07 percentage points while injecting fewer entries (Appendix C.1 and Table 10). The Local-only and Global-only ablations additionally support combining both memory sources.

Our newly added pooled-record control on ALFWorld with Qwen3-32B shows that MemCo achieves higher Held-out accuracy with fewer interaction steps under the same source experience and retrieval/composition procedure. The control pools 3,664 records from the same frozen local memories and preserves each agent’s local graph and configured budgets, while omitting global abstraction, cross-source evidence aggregation, and promotion. Full improves Held-out accuracy by about 3.2 points using the original reference and 2.50 points in the separate reevaluation. These results support the combined contribution of global abstraction, cross-source evidence aggregation, and promotion to Held-out performance beyond directly pooling the same local experience. Pooled records achieve higher Familiar accuracy, so the observed advantage concerns Held-out Layouts.

We will make the shared training-task setting, each method’s trajectory provenance and memory-construction procedure, and the acquisition, environment-interface, and few-shot settings explicit in Section 4 and Appendix F.

SettingFamiliar Acc. (%)Familiar StepsHeld-out Acc. (%)Held-out Steps
Pooled records98.0510.3794.6412.28
Full: original reference97.409.7497.8611.42
Full: separate reevaluation96.4310.4797.1411.69
AuthorsOpenReview ↗

Response to the Reviewer (2/2)

W2. How are record induction, normalization, phase identification, role mapping, and informativeness filtering implemented?

A2. The submitted version provides algorithms for local memory construction (Algorithm 1), and global promotion (Algorithm 2) with the specified implementation details of record induction, normalization, phase identification, role mapping, and informativeness filtering in Appendix B. We clarify each operation below with its algorithmic basis and a benchmark implementation example.

Normalization and phase identification. Algorithm 1 (steps 6–8) defines normalized state–action–feedback records. Appendix B.5 specifies the normalized fields supplied by environment interfaces. During memory construction, phases are identified from available interaction, progress, and completion signals. For example, during PDDL memory construction, state and goal conditions are lowercased and their separators standardized. Before termination, zero satisfied goals gives initial_planning at the initial step and search_preconditions later.

Workflow and precondition induction. Algorithm 1 (steps 11, 15–18) and Appendix B.2.3 specify extracting workflows and empirical precondition patterns in observed episode order. Workflow patterns capture ordered progress phases and associated operations, while empirical precondition patterns capture the observed context and transition associated with successful actions. For example, PDDL and ScienceWorld extract precondition records from successful actions without failure signals. Workflow records additionally require a phase change, state change, or termination. Figure A.2 illustrates the corresponding “prefer open boot” condition and phase transition.

Role mapping. Algorithm 5 (steps 7–9, Appendix B.3.1) preserves applicability conditions, action order, and outcomes while replacing entities with supported task roles. Entities without a supported role mapping remain applicability constraints. For example, ALFWorld uses object, destination, and tool fields from task descriptions to identify roles, removes instance numbers, and retains normalized entity names when no supported mapping exists.

Informativeness and sharing. Appendix B.3.2 and Table 5 specify the required content for each record type. Algorithm 2 (steps 10–14) combines structural and applicable low-information checks with Wilson promotion using aggregated evidence. Structural checks assess whether the required transferable content is present. For example, a workflow requires a condition or task role, an ordered sequence, and an outcome. The separate low-information check excludes generic advice without a usable condition, relation, or outcome. As a benchmark-specific informativeness check, PDDL filters generic “check valid actions” rules before sharing. Evidence signs are assigned from normalized interaction signals without an additional LLM call (Appendix B.2.5). Records passing these checks also satisfy the Wilson score using aggregated supporting and contradicting episode-level evidence before promotion.

The corresponding code is provided in Abstract of our paper. We involve the operational details and benchmark-specific examples alongside the relevant algorithms in Appendix B.1.

Dear Reviewer,

Thank you again for your constructive suggestions.

As the end of the author–reviewer discussion period is approaching, we would welcome any further feedback on our responses and revised manuscript. We would be happy to provide any additional information you may find helpful.

Thank you very much for your time and consideration.

The Authors

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

The paper introduces the MemCo framework for LLM-based agentic systems, which automatically manages memory from past rollouts (interactions) to help the system in future exercises be more efficient and reach higher performance. Specifically, the MemCo module comprises of two major parts: (1) a local memory that belongs to each individual agent, (2) a global memory shared across all agents. The crux of the method revolves around how information gets elevated from local to global memory so that the system remains efficient and agents only share pertinent generalizable information with other agents.

The experiments are conducted on agentic benchmarks that involve manipulation within a simulated environment, and the results show that MemCo beats previous memory management baselines both in terms of final agent accuracy and in terms of efficiency (number of steps / completion), indicating a gain in efficiency (and reducing rediscovery).

Critical strengths
  1. The paper in generally interesting, and the topic of management of memory in a hierarchical way seems fairly timeline.
  2. The experimental results are convincing, and involve the comparison of multiple recent baselines on several benchmarks, with the gains presented seeming significant.
Critical weaknesses

See questions to the authors below.

Questions for authors
  1. The paper's framing around "multi-agent" operation is a bit of a stretch. Multi-agent operation typically is concurrent (multiple agents operating at the same time, and having to coordinate operation). Yet, I believe that in this work, each experiment involves at most the operation of one agent. Q1: Do any of the benchmarks involve multi-agent operation that involves multiple agents operating concurrently in a given environment? If not, I would reframe the work to talk about specific and generic memories, but not talk about Agent_{i}, as there's really a single agent, and the question is around whether memory should be stored in a memory store specific to an environment (a room or task), or a generic/global memory. The method can remain the same, but the framing would need to change, as it is otherwise confusing.
  2. The literature on memory optimization seems highly related to the literature of tool optimization for agents (for example: https://arxiv.org/abs/2504.06821). Q2: Can the authors discuss whether their method could be applied to tool optimization, and what the key differences between tool and memory optimizations are?
  3. Can the authors discuss qualitatively the kinds of information in an example environment that get correctly/incorrectly committed to global memory (from local memory). The example in Figure 2 ("the pot is dirty") is incomplete and does not convey the method. It is unclear to me the kind of information that fits global memory that isn't "world knowledge" but is still more general than local memory information. In short, a clear example of the local memory | global memory | world knowledge hierarchy would be very useful in understanding this work.
  4. Memories tend to accumulate and make systems slower over time. Do the authors have thoughts around how to prune/consolidate memories over time to decrease the impact of stale/unused memories in the system as it operates in long horizon environments?
4 comments
AuthorsOpenReview ↗

Response to the Reviewer (1/3)

We thank the reviewer for the thoughtful assessment and four constructive questions. In particular, your question about the multi-agent setting helped us clarify how independently acquired experience is shared and reused among agents.

Q1. Multi-agent framing and memory-centric collaboration. Do any benchmarks involve multiple agents operating concurrently in the same environment? If not, should the work be reframed in terms of environment-specific and generic/global memories, without emphasizing agent indices?

A1: There might be some confusion. We appreciate the opportunity to clarify our setting. MemCo focuses on memory-centric collaboration among independently acting agents. Based on this motivation, our evaluation simulates the setting without concurrently operating multiple agents. Our setting cannot be simply reframed as a single agent maintaining environment-specific and generic memories. Because such proposed reframing overlooks an essential distinction: which agents acquired the experience and which agents can access it.

The central question is how experience independently acquired by different agents can be consolidated into shared knowledge and subsequently benefit other agents, including when they encounter environments in which they have no prior experience. This distinction also explains why the agent index (ii) is essential to our problem formulation (Section 3.1). The index identifies the agent that acquired the experience and determines its locally available memory Mil\mathcal{M}_i^l. Crucially, whether an environment is seen or unseen is defined relative to each agent's interaction history, not as an intrinsic property of the environment. For example, the same environment may be familiar to Agentj\text{Agent}_j but unseen to Agenti\text{Agent}_i. Agenti\text{Agent}_i can draw on transferable knowledge contributed by Agentj\text{Agent}_j despite lacking its own experience there. This agent-relative distinction would be lost if memories were indexed solely by environment, unless agent-specific experience ownership and accessibility were explicitly reintroduced.

MemCo operationalizes this setting by maintaining separate local memories Mil\mathcal{M}_i^l, aggregating evidence across agents to construct shared global memory Mg\mathcal{M}^g, and allowing each agent to combine its own local experience with transferable knowledge contributed by others during task execution.

This distinction is also reflected in our evaluation settings. Each agent is evaluated independently on the common merged test set using its own frozen local memory and the shared frozen global memory. The test set includes tasks from both its acquisition domain and other domains, allowing the agent to draw on experience contributed by others (Appendix F.5). The memory-scope ablation supports the benefit of this sharing: on ALFWorld with Qwen3-32B, Full MemCo achieves 97.86% accuracy on Held-out Layouts, compared with 89.64% for Local-only, using the same frozen local-memory (Table 10).

AuthorsOpenReview ↗

Response to the Reviewer (2/3)

Q2. Memory optimization appears closely related to tool optimization. What are the key differences, and could MemCo be applied to tool optimization?

A2. Thank you for the question. Memory optimization and tool optimization are closely related, but differ in their primary optimization targets. Memory optimization focuses on how experience is organized, retained, and retrieved to guide decisions. Tool optimization concerns the tools or skills through which agents act and how they are used, including constructing or refining callable procedures. Both can draw on interaction experience, and reusable procedures stored in memory could serve as candidates for tool construction.

The cited Agent Skill Induction (ASI) work exemplifies executable skill construction. MemCo and ASI both extract reusable experience from interaction trajectories, but differ in their representation, validation, and use. MemCo stores structured records, including workflows, preconditions, and repairs. It uses aggregated supporting and contradicting evidence to select global records, then grounds retrieved records in the current state and available actions for decision-making. ASI induces executable Python programs that compose existing actions. It tests them by executing tasks with rewritten trajectory prefixes, checking task success, skill use, and environment changes before adding accepted skills to its action space. Thus, MemCo provides decision guidance, while ASI makes induced procedures directly callable.

MemCo also supports agents using existing tools on FEVER. Agents use Search and Lookup to retrieve evidence before making a verification decision. Table 1 reports the highest mean accuracy and fewest mean interaction steps for MemCo across all three main backbones. With GPT-4o-mini, accuracy improves from 63.00% with ICL to 70.90%, while steps decrease from 6.33 to 2.63. These results demonstrate MemCo's effectiveness in a tool-using setting.

Our related-work discussion covers reusable skills and procedural memory, including Voyager and SkillGen. We will add ASI and clarify the relationship between memory optimization and executable skill induction.

Q3. What information should or should not be committed to global memory, and how does it differ from local experience and general world knowledge?

A3. In MemCo, the selected pattern records of workflow, action preconditions, repairs, and failures, as well as actionable knowledge records to be shared for global memory. However, fine-grained trajectories, local memory graphs, and instance-specific reflections should not be shared for global memory, as these are less transferable. All shareable information is abstracted, and checked by Wilson score, only records satisfying the threshold can be promoted as global memory. (Sections 3.3–3.4 and Appendix B.3). For example, removing the cleanable-target condition or making a particular kitchen sink a universal requirement would be an incorrect abstraction.

Local memory retains an agent’s own task experience and execution context. Global memory stores selected records for reuse across agents under their applicability conditions. World knowledge refers here to general background facts and is included only for comparison.

As requested, we highlight the schematic clean-before-place example in Table 6, alongside the corresponding local workflow in Table 13 in the original submission, now Table 17 in the revised manuscript.:

InformationConcrete example
Local memoryFind a basin, clean the soap, and place it in the vanity. A local precondition records that cleaning advances the cleaning phase when dirty soap is held at a usable basin.
Global memory: workflowCondition: a cleanable target. Actions: find a water source, clean the target, and place it at the goal destination. Outcome: a stored clean target.
Global memory: preconditionWith a dirty target held at a usable cleaning tool, cleaning advances the corresponding phase.
World knowledge, for comparison“Water can be used for cleaning” is a general fact without the above recorded procedure or its supporting interaction evidence.

In this example, the local workflow refers to soap, a basin, and a vanity. During construction, MemCo abstracts these entities into task roles, e.g., target object, water source, and goal destination, and promotes the supported abstract record to global memory, preserving its applicability condition, action order, and outcome (Section 3.4). During use, MemCo grounds the retrieved workflow in the current environment, binding its roles to the corresponding objects and checking compatibility with the current state and available actions (Section 3.5). The world-knowledge statement identifies a general use of water but does not specify this conditional procedure or its outcome.

Figure 2 shows “the pot is dirty” as the current observation.

AuthorsOpenReview ↗

Response to the Reviewer (3/3)

Q4. How can MemCo consolidate or prune accumulating memories to reduce the impact of stale or unused records during long-horizon operation?

A4. To reduce the impact of stale or unused memories accumulated over time, MemCo consolidates repeated experience and rebuilds global memory using records that satisfy its current promotion criteria. Query-guided retrieval and composition then select support relevant to the current state, task phase, and goal to reduce interference from irrelevant or stale memories during decision-making.

During memory construction, equivalent local graph elements and records are merged with accumulated evidence. Global memory maintains one record per abstract key and excludes records that fail structural, low-information, or Wilson checks when rebuilt (Sections 3.3–3.4 and Appendix B.2–B.3). Records are not deleted solely because they are old or unused. The global-promotion ablation shows that Wilson retains a smaller bank with higher accuracy than empirical-support promotion. On ALFWorld with Qwen3-32B, using the same frozen local-memory snapshot, Wilson retains 426 rather than 736 records and improves Familiar and Held-out accuracy by 2.27 and 1.43 percentage points, respectively (Table 10).

During decision-making, retrieval selects records by state, task-phase, and goal relevance, reducing the interference of irrelevant memory. Incompatible records are excluded, stale local information receives lower priority, and composition grounds shared guidance in the current environment using available local evidence (Section 3.5). The corresponding ablation shows higher accuracy with less injected context from the same stored memories. Compared with Fixed L5+G5, Full MemCo reduces average injected entries from 9.99 to 1.70 and memory-prompt tokens from 207.56 to 123.23, while improving Familiar and Held-out accuracy by 10.71 and 16.07 percentage points (Table 10).

We additionally measured offline retrieval overhead as the shared memory bank grows. Starting from the same pool of 749 candidates passing structural and informativeness checks, before Wilson thresholding, we constructed nested test banks containing 25%, 50%, 75%, and 100% of the pool. We kept each agent’s local memory and the same recorded queries fixed, repeating each query three times at every bank size. As the bank grew from 25% to 100% records, median retrieval time increased only from 45.91 to 58.77 ms, showing that retrieval remains efficient even with four times as many shared records at the tested scales.

Fraction of eligible poolShared recordsMean retrieval (ms)Median (ms)p95 (ms)
25%18746.7445.9162.59
50%37552.0150.4366.90
75%56256.7954.6573.02
100%74961.0258.7776.91

Dear Reviewer,

Thank you again for your constructive suggestions.

As the end of the author–reviewer discussion period is approaching, we would welcome any further feedback on our responses and revised manuscript. We would be happy to provide any additional information you may find helpful.

Thank you very much for your time and consideration.

The Authors

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

Thank you for your questions and suggestions. We summarize the main clarifications in our replies. MemCo studies how experience acquired by one agent can help another agent act in an unfamiliar environment. Our additional controlled experiments show where constructing transferable global memory improves task performance beyond directly pooling the same experience.

  1. Memory-centric collaboration among independently acting agents. Each agent builds local memory from its own task executions. Other agents can retrieve selected records through shared global memory and ground them in their current environment. In the clean-before-place example from our response to Reviewer axDr’s Q3, the basin and vanity are specific to the local experience. The shared workflow specifies that, for a cleanable target, the agent should find a water source, clean the target, and then place it at the goal destination.

  2. Acquisition conditions and implementation details. Within each benchmark, all memory-based methods receive the same training tasks and maximum environment-interaction budget. Each method builds memory from its own trajectories following its original methods. The revised Appendix B.1 gives benchmark examples of how MemCo constructs, abstracts, and filters records. The code is linked in the last sentence of the abstract.

  3. Performance gains beyond pooling the same experience. Full MemCo and the pooled-record control use the same source experience and retrieval and composition procedure. On ALFWorld with Qwen3-32B, the largest gains occur on other-domain tasks under Held-out Layouts, where MemCo achieves 98.57% accuracy, compared with 94.29% for pooled records and 86.67% for Local-only (revised Table 12). With the same stored memories, query-guided retrieval and composition achieve higher accuracy than fixed concatenation while injecting fewer entries (Table 10). At an equal count of 426 records, the Wilson reference also exceeds the mean accuracy of three empirical-support selections on both layout splits (revised Table 13).

After reviewing the additional controlled experiments, Reviewer ReQP wrote, “These results adequately address my concerns.” We appreciate your consideration of these clarifications and results in the final assessment.

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.