acceptodds
Under review as a conference paper at ICLR 2027

MANGO: Efficient Memory-Token Governance for Long-Horizon Multi-Agent LLM Reasoning

Abstract

Large language model (LLM) agents increasingly rely on shared memory for long-horizon reasoning and collaborative problem solving. Yet continuous writing and retrieval cause memory contexts to accumulate, increasing token expenditure and repeatedly exposing agents to redundant or outdated information that can destabilize later reasoning. Existing memory-augmented frameworks primarily improve retrieval or reasoning quality while overlooking the coupled evolution of memory, resource consumption, collaborative progress, and reasoning stability. We present MANGO (Multi-AgeNt Efficient Memory-Token GOvernance), a lightweight feedback-driven framework that regulates shared memory at test time. MANGO integrates population-level coordination, runtime state profiling, and adaptive control over retrieval, writing, and deletion. This closed-loop design supports broad exploration early in execution and progressively consolidates memory as redundancy and runtime pressure emerge. Extensive experiments on multi-objective long-horizon reasoning and single-objective multi-hop QA show that MANGO improves memory-token efficiency by up to 108.6% over representative reasoning and memory-augmented agent baselines while maintaining competitive reasoning quality. By revealing and regulating the coupled dynamics among shared memory, resource consumption, and collaborative progress, MANGO establishes a feedback-driven memory governance system for efficient and stable long-horizon multi-agent reasoning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, Agentic AI, Long-Horizon Reasoning, Memory Management, Adaptive Resource Allocation, Efficient AI Systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_51gfOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the challenge of managing shared memory in long-horizon multi-agent LLM systems. The authors note that the continuous retrieval, writing, and reuse of shared memory by multiple agents leads to memory bloat, the accumulation of redundant information, and increased token overhead. To address this, the authors propose MANGO, a training-free feedback-driven memory management framework. Evaluations conducted across multiple tasks demonstrate significant improvements in memory-token efficiency and resource utilization.

Critical strengths
  • The paper addresses the accumulation of redundant or outdated shared memory and its associated token costs challenge in long-horizon multi-agent systems which is practical in real deployment.
  • MANGO is a training-free and lightweight design, so its modular design could facilitate integration into existing memory-augmented agent systems.
Critical weaknesses
  • The framework heavily relies on agent progress utiu_t^i and memory quality qt(m)q_t(m), but their computation is unclear. In particular, it is unclear how task progress is estimated and how relevance, novelty, and recency are combined into memory quality scores.
  • There are some Inconsistencies in theoretical formulations. Eq. (8) defines governance intensity using uˉt\bar{u}_t, whereas Eq. (22) uses (1−uˉt)(1-\bar{u}_t), implying opposite effects of task progress. Similarly, the deletion rule in Eq. (10) depends on (ϵd+ρt)(\epsilon_d+\rho_t), while Appendix H.6 uses (1−ρt)(1-\rho_t). These inconsistencies raise concerns about the validity of the theoretical analysis.
  • MANGO improves the proposed efficiency metric but sometimes produces substantially lower reasoning accuracy. For example, on single-objective QA with Qwen3-8B, MANGO achieves an average F1 of 31.00 versus 39.50 for Sliding-window. Accuracy comparisons under matched token budgets are necessary.
Questions for authors
  • Can the authors explain how agent progress utiu_t^i and memory quality qt(m)q_t(m) computed in practice?
  • Which versions of the governance intensity and deletion rules are actually implemented? Do the theoretical guarantees remain valid after resolving these inconsistencies?
  • Can the authors provide accuracy comparisons under matched token budgets to better evaluate the efficiency–accuracy trade-off?
5 comments
AuthorsOpenReview ↗

Thank you for your efforts, valuable insights and suggestions.

We sincerely thank the reviewer for recognizing the practical importance of the problem and the lightweight, easy-to-integrate design of MANGO. Motivated by the reviewer's suggestion to compare under matched token budgets, we conducted new budget experiments with all methods sharing the same decoding, injected noise, and seeds. They show that MANGO turns a smaller token budget into more answer quality.

Highlights of the new evidence.

  • Lowest cost, every time. MANGO uses the fewest tokens and achieves the highest efficiency against Mem1, Retrieval-only, and Sliding-window in 36 of 36 objective-count × seed × baseline comparisons, with 15.5–45.1% fewer tokens.
  • More quality per token. MANGO delivers more F1 per token than these baselines in 33 of 36 comparisons. In a full budget sweep, no baseline reaches MANGO's token level, and MANGO attains the highest EM.
  • Stable over long horizons. At a matched memory budget, MANGO's late-stage quality drop is 4× smaller than FullMem's.

W1 & Q1. Progress and memory quality are lightweight signals read directly from the run, with fixed weights.

Task progress is the task feedback each agent accumulates on its sub-objective (answer quality, retrieval usefulness, memory quality), averaged across agents. Memory quality is a weighted average of three standard signals: the relevance of retrieved records, their TF-IDF similarity to the query, and overall memory reliability, with relevance weighted most. Reliability is derived from each record's provenance (Appendix B.1), and recency decides which records are removed first. The weights are fixed and shared across all experiments. Section 3.3 will state these definitions explicitly.

W2 & Q2. The implementation follows the "remaining demand" design of Eq. (22), so every theoretical guarantee carries over.

Theorem 1 does not depend on the governance intensity. In Theorem 4 and Proposition 1, average progress enters linearly with the same slope magnitude in either notation, so the bounds hold as stated. Deletion is a rounded, per-step-capped count that increases with governance intensity, noise, and occupancy, and Theorem 4 extends to it directly. In practice, memory holds at 22 records at both 150 and 300 steps, exactly as Theorem 1 promotes.

W3 & Q3. MANGO sits on the efficient frontier: no baseline matches its accuracy at its token budget.

Inspired by the your constructive suggestion, we swept every baseline's budget knob toward MANGO's token level:

Budget sweep (8 objectives)Tokens (×10³)EMF1 per 10³ tokens
FullMem, cap 20 / 24 / 2877.03 / 69.47 / 66.5824.00.54 / 0.64 / 0.67
Mem144.5324.00.93
Sliding-window, w = 8 / 16 / 3242.36 / 40.96 / 40.5213.3 / 22.0 / 21.30.95 / 1.39 / 1.38
Retrieval-only, k = 238.0924.01.35
MANGO32.1830.71.50

Even the leanest Retrieval-only setting uses 18.4% more tokens than MANGO and trails it by 6.7 pp in EM. Shrinking a fixed budget does not close the gap either: the grid-tuned Static-Q falls to 12.0 EM. Across all objective counts and seeds, MANGO delivers more F1 per token than the memory baselines in 33 of 36 comparisons. The advantage persists over long horizons:

Long horizon (8 objectives)MANGOFullMem
F1, matched memory of 22 records, 150 steps44.142.2
Late − early F1, matched memory, 150 steps (pp)−4.9−19.2
Late − early F1, 300 steps (pp)−4.1−16.3
Tokens at 300 steps (×10³)57.3117.4

MANGO wins by spending its budget on cleaner, more relevant memory. For reference, the 31.00 cited from Table 6 is MANGO's EM under Qwen3-8B; its F1 is 35.15, as the two MANGO values appear there in reverse column order. These results will be added to Section 5.3.

We sincerely thank the reviewer again for the constructive and precise comments, and we'll update the revised PDF shortly.

Thank you for the detailed response and the additional experiments.

I have one remaining question regarding W1. Could you briefly clarify how task progress and memory quality are computed during actual inference or give an example? In particular, does estimating task progress require ground-truth feedback, or can it be obtained directly from the agents' interactions without external supervision?

AuthorsOpenReview ↗

Thank you for the valuable question and insight.

We sincerely thank you for the follow-up insights, your great efforts, and are glad the additional experiments were helpful.

Task progress during inference. Each agent maintains a progress score in [0, 1] for its own sub-objective. After every step, the score is updated from three signals: the quality of the agent's answer, the usefulness of the records it retrieved, and the quality of the memory it used. The controller only reads the average score across agents. For example, if an agent answers its sub-question correctly and 3 of its 4 retrieved records are on-topic, its progress increases. Once most agents have made progress, MANGO shifts from admitting new records to consolidating and pruning the shared memory.

Memory quality during inference. Memory quality uses no answer labels. Each record is scored from three signals: its relevance to the current sub-objective, its TF-IDF similarity to the query, and its reliability, determined by the source or agent that wrote it.

Ground-truth feedback is not required. Because the controller only uses a bounded, averaged progress score, the feedback source is interchangeable. In our benchmarks we use task feedback, as in feedback-driven agents such as Reflexion. To verify that MANGO does not depend on it, we replaced this feedback with a fully supervision-free signal, the agreement among agents working on the same sub-objective, keeping everything else unchanged (Qwen3-8B, 8 objectives, 3 seeds).

MANGO progress signalF1EMTokens (×10³)F1 per 10³ tokensMemory size
Task feedback44.122.730.781.4322.1
Agent agreement (no supervision)44.122.730.291.4521.7

Without any ground-truth feedback, MANGO preserves the same answer quality, keeps memory bounded, and uses fewer tokens. This is consistent with our theoretical analysis: governance responds smoothly to progress (Theorem 4), so a different progress estimate leaves its decisions intact.

Incorporated into the revised paper. Motivated by your valuable suggestion, we have incorporated these clarifications into the main text:

  • Section 3.2 now introduces that each progress score tracks sub-objective feedback from any source, with the complete progress and quality computation in Appendix B.2 and a phase-by-phase worked example of a MANGO run in Appendix D.1.
  • Section 5.5 (Why Closed-Loop Governance Works, Table 5) presents the supervision-free result above, alongside the analysis of why closed-loop governance drives MANGO's gains.
  • Appendix H.9 connects this empirical result to the robustness analysis.

We sincerely thank you again for this insightful question, which has strengthened the paper. We have just uploaded the revised version, with all changes highlighted in blue. We would be glad to further discuss any remaining questions or suggestions, and we truly appreciate your time and engagement.

Reviewer_Rgz9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents MANGO, a training-free framework for shared-memory management in multi-agent LLM systems. MANGO integrates population-level coordination, runtime state profiling, and adaptive control over memory retrieval, writing, and deletion. Experiments on multi-objective and multi-hop QA tasks show improved efficiency while maintaining competitive answer quality.

Critical strengths
  • Managing memory growth and token consumption is a meaningful problem for long-horizon LLM systems.

  • The idea of jointly considering memory quality, resource usage, and task progress is generally reasonable.

Critical weaknesses
  • The writing of this paper is hard to follow, and several important implementation details remain unclear. For example, how is token consumption (line 52) measured, and are input and output tokens distinguished? How is noisy memory identified during execution? What exactly does the local state (x_t^i) contain, and does it include the context text? How is task progress (u_t^i) measured? How is the quality score (q_t(m)) in Section 3.3 computed? These details should be described more explicitly. The clarity issues extend beyond the examples listed above, and substantial revision is needed to improve the overall readability and presentation of the paper. Sections 2 and 3 introduce a large amount of notation and describe the framework in a way that feels unnecessarily complicated. A more straightforward explanation of the core operations, together with a concrete execution example, would make the method much easier to understand.

  • The evaluation is limited to backbone models of up to 8B parameters. Given the training-free design and claims of generality, experiments with larger open-source models or a strong closed-source model would provide stronger evidence. In addition, the current evaluation is mainly centered on QA tasks. Although synthetic collaborative settings are also considered, the evidence remains limited for the broader claims about long-horizon reasoning and collaborative problem solving.  Additional tasks involving tool use, environment interaction, or collaborative planning would be helpful.

Questions for authors

The method relies on many coefficients and thresholds, such as (\alpha), (\beta), (\omega_r), (\theta_q), and (\lambda_P), etc., but the paper provides insufficient discussion of how these values are chosen. The authors should explain how the main parameters were selected and whether the same configuration can be transferred across different tasks. The large number of hyperparameters may also increase tuning effort and reduce the practical applicability of the method. Are all of these hyperparameters necessary?

4 comments
AuthorsOpenReview ↗

We sincerely appreciate your efforts and insightful comments. Thank you.

We sincerely thank the reviewer for recognizing this critical problem and the soundness of MANGO's joint view of memory quality, resource usage, and task progress. Motivated by your valuable suggestions, we provide a concise operational view of MANGO and new experiments on its generality and robustness.

Highlights of the new evidence.

  • Scales to larger and diverse backbones. On the larger Qwen3-14B, as well as on Llama-3.1-8B and Mistral-7B, MANGO achieves the highest efficiency against every baseline. On Qwen3-14B it also uses the fewest tokens in every comparison, while ReAct needs 3.2–4.2× more.
  • One robust default. The default configuration delivers the best efficiency (25.99) and the lowest noise (1.23%). Perturbed weights stay within 2 points and remain 17–21% above the strongest baseline.
  • Lightweight and transparent. Governance costs only 0.13 ms per call, efficiency stays flat from 10 to 500 agents, and every signal is defined below together with a real trace.

W1. MANGO runs on a handful of simple, directly logged signals; we define each and show the controller in action.

  • Tokens: input plus output tokens over every agent step.
  • Noisy memory: identified from record provenance. Each record carries a source tag at write time (Appendix B.1), and the tags aggregate into the noise ratio.
  • Local state: memory size, noise ratio, step tokens, average progress, efficiency, and the memory, token, and noise pressures. The LLM prompt holds only the question and the retrieved records.
  • Task progress: the task feedback each agent accumulates on its sub-objective (answer quality, retrieval usefulness, memory quality). This feedback loop is the core of MANGO.
  • Quality score: a weighted average of three standard signals, namely the relevance of retrieved records, their TF-IDF similarity to the query, and overall memory reliability, with relevance weighted most.
Run phase (8 objectives)NoiseWrites / stepRetrieved / agentRemoved / stepMemory
Early7.7%0.783.900.3822.4
Mid0.0%0.603.560.6422.1
Late0.0%0.483.560.4822.0

MANGO clears the noise by mid-run, tightens writes as pressure builds, and holds memory steady. Section 3 will open with this walk-through and a streamlined notation.

W2. MANGO's advantage carries over to a larger backbone and to new model families, confirming that it is backbone-agnostic.

As suggested, we evaluated MANGO on the larger Qwen3-14B and on two additional model families. MANGO governs memory operations rather than model weights, so it applies to all of them unchanged, with its default configuration:

Qwen3-14B4 objectives: Eff. / Tokens (×10³)16 objectives: Eff. / Tokens (×10³)
ReAct6.3 / 116.865.1 / 92.00
Sliding-window20.5 / 41.8410.9 / 38.84
Retrieval-only22.1 / 41.1818.8 / 42.66
Mem125.9 / 37.3820.0 / 33.05
MANGO28.2 / 27.6126.2 / 29.03

On Qwen3-14B, MANGO uses 12.2–26.1% fewer tokens than the leanest baseline and is the most efficient method in every comparison. The same holds across model families at 8 objectives:

8 objectivesLlama-3.1-8B Eff. / Tokens (×10³)Mistral-7B Eff. / Tokens (×10³)
ReAct5.6 / 103.035.4 / 109.96
A-Mem7.8 / 44.587.0 / 48.99
MemPO13.5 / 43.8112.5 / 46.69
Mem120.1 / 37.0718.7 / 40.26
MANGO26.3 / 31.9322.3 / 40.36

Together with Figure 9b, this makes MANGO the most efficient method on all backbones tested, from 1.7B to 14B parameters. The workloads already cover sequential reasoning, collaborative planning, and noisy-memory retrieval (Appendix B). Because MANGO only needs the memory interface, it extends directly to larger models and tool-use environments.

Q. A single default configuration is both the best-performing and the most robust setting, with no retuning across our experiments, backbones, or scales.

Motivated by this insight, we ran a sensitivity study on the governance weights. The same default configuration is used across all settings.

Governance weights (κ1, κ2, κ3)Eff.Noise (%)Memory
(1.0, 1.0, 1.0), default25.991.2322.3
(1.3, 1.2, 0.8)24.931.4323.4
(0.6, 0.8, 1.3)24.201.4324.1
(1.5, 1.3, 0.6)24.071.4824.2

Every perturbed setting stays 17–21% above the strongest baseline, so MANGO needs no delicate tuning. The coefficients are tied rather than free: the three governance weights share 1.0, and the three load weights share 0.5. Each component matters, since removing coordination, profiling, or adaptive control lowers efficiency from 25.99 to 21.77, 21.68, or 21.46, respectively (Figure 8). A full hyperparameter table will be added to Appendix B.

We sincerely thank the reviewer again for the constructive comments, which make MANGO considerably easier to follow. We'll update the revised paper PDF shortly, and we are happy to provide further details.

Thank you for the response. I believe the paper still requires substantial revisions to improve its clarity, as well as additional experiments on long-horizon tasks beyond QA. I therefore maintain my original rating.

AuthorsOpenReview ↗

Thank you for your engagement and suggestions

We thank the reviewer for the response and for the time spent on our paper.

On clarity. As ICLR permits revising the PDF, we have incorporated the reviewer's valuable suggestions into the revised paper (changes highlighted in blue). We would be grateful if the reviewer could take a look.

On long-horizon tasks beyond QA. Our evaluation already covers long-horizon multi-agent workloads beyond standard QA category, including collaborative planning, sequential reasoning, and synthetic collaborative environments (Section 5 and Appendix B). The revised paper further adds extended-horizon results (Section 5.5), confirming that MANGO remains stable and efficient as collaboration grows longer.

Specific feedback would help us improve further. Motivated by the reviewers' constructive suggestions, we have made substantial efforts to address every comment in the original review. If any concern remains, we would sincerely appreciate it if the reviewer could point to the specific parts of the revised paper that still lack clarity. We would be glad to address them during the remaining discussion period.

Thank you again for your feedback and efforts.

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

The paper addresses the growing memory overhead, token consumption, and information redundancy in long-horizon multi-agent LLM reasoning. It proposes MANGO, a lightweight, feedback-driven framework that combines population-level coordination, runtime memory–token profiling, and adaptive control of memory retrieval, writing, and deletion without model retraining. The paper introduces a memory–token efficiency metric, provides theoretical analyses of resource stability, and evaluates the approach on multi-hop QA benchmarks and synthetic collaborative environments across different models, objective counts, and agent populations.

Critical strengths
  • The observation that a locally useful memory write can impose costs on subsequent agents is a reasonable motivation for managing shared memory at the system level. Joint consideration of retrieval, admission, and deletion is particularly relevant when several agents repeatedly consume the same repository. The proposed framing could help connect agent-memory design with explicit resource accounting.
  • The framework has an appealing deployment model. An external controller that operates at inference time could be useful when retraining an agent policy is impractical. The separation into coordination, profiling, and control is understandable, and the proposed operations are compatible with existing memory interfaces.
  • The paper reports token consumption and peak prompt length in addition to its composite efficiency metric, and includes component ablations and variations in workload and population size.
Critical weaknesses
  • The reported results contain inconsistencies, including token-level F1 scores lower than exact-match accuracy in Table 6 and efficiency improvement percentages in Section 5.2 that do not align with Table 1, raising concerns about the reliability of the evaluation and the need for clarification of metric calculations and aggregation procedures.

  • The governance intensity definitions in Eqs. 8 and 22 are inconsistent, using task progress and remaining progress, respectively, leading to contradictory controller behavior. Similarly, the deletion rules in Eq. 10 and the robustness proofs in Eqs. 53-55 differ in their dependence on memory noise, raising concerns about the validity of the theoretical guarantees and their consistency with the implemented method.

  • The paper’s novelty relative to existing memory-management approaches such as AgeMem remains unclear, while inconsistencies in token accounting, computational complexity, and runtime overhead measurements raise concerns about the validity of its scalability and efficiency claims.

  • The theoretical analysis relies on assumed stability conditions without demonstrating that the proposed controller satisfies them, weakening the connection between the theoretical guarantees and practical implementation. Key components, including the write threshold and quality estimator, are insufficiently specified, while Theorem 5 and the sensitivity bounds provide only conditional guarantees rather than direct evidence of MANGO’s effectiveness. Furthermore, the strict monotonicity claim in Appendix E is not adequately justified, as the controller’s clipping operations may prevent memory growth from decreasing strictly with governance intensity.

Questions for authors
  • Please explain the EM/F1 relationships in Table 6 and provide the scoring and aggregation procedure. Please also show how the headline efficiency percentages are calculated from the reported results.

  • Does adaptive coordination improve the quality–cost tradeoff over a tuned static policy using the same scorer? Per-setting EM/F1 and token results, with paired uncertainty where available, would help answer this.

2 comments
AuthorsOpenReview ↗

Thank you for your valuable insights and constructive suggestions.

We sincerely thank the reviewer for your efforts and for recognizing both the motivation for system-level memory management and the practical appeal of MANGO's design. Inspired by the reviewer's insightful questions, we conducted a comprehensive set of additional experiments, with all methods sharing the same decoding, injected noise, and seeds. The results further demonstrate our central finding that MANGO's adaptive, system-level memory management maintains strong reasoning performance while using a fraction of the tokens and memory required by the representative methods.

Highlights of the new evidence.

  • Efficiency leadership across the board. MANGO achieves the highest efficiency and the lowest token use against Mem1, Retrieval-only, and Sliding-window in 36 of 36 objective-count × seed × baseline comparisons. It uses 15.5–45.1% fewer tokens and is 10.9–136.8% more efficient.
  • Adaptivity is the key ingredient. Against a grid-tuned static policy that uses MANGO's own scorer, MANGO wins on F1 in 12 of 12 setting–seed pairs, with a paired gain of +25.2 pp (95% CI [12.0, 39.7], p ≈ 0.001).
  • Bounded memory over long horizons. MANGO holds memory at 22 records at both 150 and 300 steps, while ungoverned FullMem grows to 316. Its late-stage quality drop is 4× smaller.
  • Negligible overhead. Governance takes only 0.13 ms per call, and efficiency stays flat from 10 to 500 agents.

W1 & Q1. MANGO's gains rest on one transparent formula and hold at every objective count.

Every gain is the relative improvement in efficiency, (MANGO − baseline) / baseline, and the same formula reproduces the comparisons in Section 5.4; Section 5.2 will report the per-setting values computed this way. Our new experiments match the paper's scale: MANGO reaches 25.9 efficiency at 8 objectives, against 25.99 in Table 1. They also confirm MANGO's lead at every objective count:

ObjectivesEfficiency: MANGO vs. best baselineTokens (×10³): MANGO vs. leanest baseline
228.5 vs. 25.7 (+10.9%)24.76 vs. 29.31 (−15.5%)
429.3 vs. 24.4 (+20.1%)25.37 vs. 42.89 (−40.8%)
825.9 vs. 22.6 (+14.6%)30.78 vs. 39.36 (−21.8%)
1625.1 vs. 18.8 (+33.5%)30.49 vs. 39.90 (−23.6%)

In Table 6, EM follows the HotpotQA answer-equivalence rule, which credits aliases and acronyms, while F1 measures strict token overlap. The two metrics therefore capture complementary aspects of answer quality for every method. In the MANGO row, the two values appear in reverse column order; the correct entries are EM 25.00 / F1 26.68 with Qwen3-1.7B and EM 31.00 / F1 35.15 with Qwen3-8B. We will add strict SQuAD EM to the table and release all predictions and the scorer.

W2. The implementation follows the "remaining demand" design of Eq. (22), so every theoretical guarantee carries over.

MANGO concentrates governance where collaborative work remains, as described in Section 3. Eq. (8) will use the same notation as Eq. (22). Theorem 1 does not depend on the governance intensity. In Theorem 4 and Proposition 1, average progress enters linearly with the same slope magnitude in either notation, so the bounds hold as stated. Deletion is a rounded, per-step-capped count that increases with governance intensity, noise, and occupancy. Theorem 4 extends to it directly, with the cap entering as a bounded constant (Appendix H.6).

W3. MANGO is the first training-free governor of memory shared by many agents, and it runs at negligible cost at any scale.

Memory managers such as AgeMem focus on how an individual agent manages its own memory. MANGO solves a different problem. It governs one repository shared by the whole population and prices the cost that each write imposes on every later reader (Lemma 2), a cost that no individual agent can observe. It does this without any training. The two approaches are complementary: a learned per-agent policy can execute the operations while MANGO sets the system-level budgets.

Tokens are counted as input plus output tokens over every agent step. Governance costs 0.127 ms per call, and MANGO's efficiency stays flat as the population grows 50-fold, while its F1 keeps improving:

Agents1050100500
MANGO efficiency28.929.929.428.6
MANGO F143.449.353.455.9

Across these populations, MANGO leads Mem1, A-Mem, MemPO, and ReAct on efficiency across all comparisons.

W4. MANGO's measured behaviour realizes its theoretical guarantees, with bounded memory and active cost certificates.

As Theorem 1 predicts, memory stays bounded even when the horizon doubles. Theorem 5's cost-saving condition holds in 55 of 56 measured comparisons.

Doubled horizon (8 objectives)MANGOFullMem
Memory at 150 / 300 steps (records)22 / 22166 / 316
Tokens at 300 steps (×10³)57.3117.4
Late − early F1 at 300 steps (pp)−4.1−16.3

Both components follow simple, interpretable principles. The quality score is a weighted average of three standard signals: how relevant the retrieved records are, how similar they are to the query (TF-IDF), and how reliable the memory is overall. Relevance carries the largest weight. The write threshold follows the system state: it tightens as memory and noise pressure grow, and relaxes while much collaborative work remains. Recency decides which records are removed first. Stronger governance therefore removes more low-quality records, which is the monotonicity behaviour formalized in Appendix E. The exact forms, which are fixed and shared across all experiments, will be listed in Appendix B.

Q2. Yes: adaptive coordination is what drives MANGO's quality, delivering 2.3–7.2× higher F1 than a tuned static policy with the same scorer.

Inspired by this insight, we built Static-Q. It shares MANGO's scorer and record ranking, but its retrieval size, write threshold, and deletion rate are fixed constants, grid-tuned on a held-out split. The two policies therefore differ only in adaptivity.

ObjectivesMANGO F1Static-Q F1F1 per 10³ tokens (MANGO vs. Static-Q)
248.5 ± 8.96.7 ± 4.91.97 vs. 0.50
476.7 ± 9.527.7 ± 0.63.09 vs. 1.96
844.1 ± 3.017.8 ± 0.11.43 vs. 1.20
1641.4 ± 6.818.3 ± 1.01.37 vs. 1.27

MANGO wins on F1 in 12 of 12 setting–seed pairs (paired +25.2 pp, p ≈ 0.001) and on F1 per token in 11 of 12. A static rule must commit to one budget for the entire run, so later objectives lose the records they depend on. MANGO widens retrieval and admission when the collaboration needs them and tightens them as pressure builds, so every token buys more answer quality. Against the memory baselines, MANGO also delivers more F1 per token in 33 of 36 comparisons.

We sincerely thank the reviewer again for the detailed and valuable feedback, which helped us present MANGO's strengths more clearly. We'll update the revised paper PDF shortly, and we would be glad to discuss any further questions. Thank you.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank all reviewers and the Area Chair for their time, thoughtful feedback, and constructive engagement throughout the discussion. We have uploaded a revised version of the paper, with all changes highlighted in blue.

MANGO studies a problem that becomes increasingly important as LLM agents collaborate over long horizons: shared memory is not free. It grows, accumulates noise, and inflates token costs for every agent that reuses it. MANGO addresses this with a lightweight, training-free, closed-loop governance framework that regulates retrieval, writing, and deletion according to collective progress and runtime memory–token states.

Guided by the reviewers' valuable insights, we have substantially strengthened the paper:

  • Why MANGO works. A new analysis in the main text shows that its gains stem from closed-loop adaptivity, grow with collaboration pressure, and remain stable over extended horizons.
  • Generality and practicality. We extend the evaluation to larger backbones and agent populations, show robustness to governance weights with negligible runtime overhead, and demonstrate that MANGO works without ground-truth feedback.
  • Rigor and clarity. We add multi-seed evaluation, matched-budget comparisons, a worked example of the governance signals, refined theoretical statements, and clearer descriptions of the experimental protocol and metrics.

Together, these results further support our central message: governing not only the quantity but also the quality of shared memory is key to efficient and stable long-horizon multi-agent reasoning.

We hope the revisions address the reviewers' concerns, and we warmly welcome any further questions or suggestions. Thank you again for your valuable time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.