acceptodds
Under review as a conference paper at ICLR 2027

The Memory Curse: How Expanded Recall Erodes Cooperation in LLM Agents

Abstract

Expanding the interaction history available to a large language model (LLM) agent is often treated as a straightforward capability upgrade, but we find that it can reduce cooperation in repeated social dilemmas. Across 7 LLMs and 4 games with interactions evaluated over up to 500 rounds, many model–game settings exhibit lower cooperation under expanded recall, a pattern we term the *memory curse*. Complementary diagnostics and interventions connect this effect to action persistence, memory content, and expressed reasoning patterns. Lexical analysis associates longer histories with less forward-looking language, supported by a blinded contextual evaluation of 2,016 rationales. An action-balanced, rationale-only low-rank adaptation (LoRA) model trained on 1,000 Public Goods traces shows a substantially smaller short-to-long-memory cooperation decline in the training game and positive average zero-shot changes in other, untrained games. Compact recent, summary, and retrieval views yield higher average cooperation than the full 80-round record; token-matched compact raw histories show similar aggregate gains to the summary and retrieval views. Removing explicit Chain-of-Thought attenuates the mean decline in most settings exhibiting the memory curse with explicit reasoning, with stability responses varying across settings. Together, these results establish accessible history as a central behavioral design choice for sustaining cooperation.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Multi-Agent Systems, Cooperative AI, Social Dilemmas, Bounded Rationality

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_3UraOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper studies how accessible interaction history affects cooperation among LLM agents in repeated social dilemmas. Across seven LLMs, four games, and multiple history lengths, the authors find that longer histories often reduce cooperation, a phenomenon termed the "Memory Curse." The paper further investigates this effect through action-transition analysis, reasoning-trace analysis, asymmetric-memory experiments, memory sanitization, removal of explicit reasoning, alternative memory representations, and rationale-only fine-tuning. The overall evidence suggests that expanded recall can increase persistence following non-cooperative actions and that both the content and representation of interaction history substantially affect long-run agent behavior.

Critical strengths
  1. Clear and well-organized presentation.
    The paper is generally easy to follow, with the experimental setup, main observations, and follow-up analyses presented in a structured manner.

  2. Broad experimental coverage.
    The study evaluates multiple LLMs, four repeated social dilemma games, and a wide range of history lengths, providing a relatively large empirical study of memory effects in repeated LLM interactions.

  3. Effective figures and tables.
    The figures clearly illustrate the main cooperation trends across history lengths, and the tables provide useful breakdowns across models, games, and experimental conditions.

Critical weaknesses

W1. A decline in cooperation does not necessarily imply a harmful "memory curse."

The central empirical finding that longer accessible histories often reduce cooperation is interesting and well supported. However, I am not fully convinced that lower cooperation should itself be interpreted as a dysfunction. In repeated games, past actions provide information about an opponent's strategy or type, and conditioning future behavior on this history can be strategically rational [1]. A long-memory agent that becomes less cooperative after observing repeated defections may therefore be adapting to evidence about its opponent rather than being harmed by additional memory.

The paper does report realized and discounted rewards in addition to cooperation. However, the definition of the Memory Cursed regime itself is still based on the decrease in cooperation from short to long memory, rather than on strategic suboptimality. This distinction matters because cooperation is not always aligned with payoff: the paper's own scripted Always-Defect control shows a setting where higher cooperation leads to substantially lower reward. I therefore think that the "curse" interpretation would be stronger if it were tied more directly to a strategy-sensitive criterion such as regret, best-response quality, or opponent-conditioned expected return.

W2. The self-play design makes it difficult to distinguish a memory-induced failure from an endogenous shift in the repeated-game regime.

In the primary experiments, agents instantiated from the same backbone repeatedly interact with one another, and the memory condition is shared across the interacting agents. Consequently, increasing history length does not only change the information available to a focal agent; it simultaneously changes the behavior of its opponents. A defection can trigger retaliation, which becomes new evidence for the other agents, inducing further policy changes and potentially moving the entire interaction into a different low-cooperation regime.

This is important from a repeated-game perspective. Classical results such as the Folk Theorem show that repeated interactions can sustain multiple history-contingent equilibrium outcomes rather than a unique cooperative regime [2]. Thus, observing that long-memory self-play converges to lower cooperation does not by itself establish that the agents are behaving less rationally or that memory is producing an individual-level dysfunction. A cleaner identification would vary the focal agent's memory while holding the opponent policy fixed.

W3. The paper lacks a controlled memory evaluation in which the long-memory policy shift can be identified as clearly maladaptive.

Because the four main environments are repeated social dilemmas, history-dependent policy changes are an expected part of strategic play. The paper does include scripted Always-Defect, Tit-for-Tat, and One-error Tit-for-Tat opponents, but these experiments are used to study the relationship between prompting, cooperation, and realized payoff rather than to compare short- and long-memory agents against a fixed opponent policy.

A more diagnostic evaluation would use scripted opponents with known and possibly time-varying policies while explicitly varying memory length. For example, an opponent could defect early and then permanently switch to cooperation, or follow a stationary stochastic policy with occasional noisy defections. In such settings, one could test whether a long-memory agent appropriately discounts stale evidence or continues retaliating after the old negative history is no longer predictive. Persistent defection that lowers reward relative to an oracle or best-response policy would provide much stronger evidence for a genuinely maladaptive memory effect.

These are not minor concerns like no repeated seeds. So my initial thought is rejection, I will consider the response from the rebuttal phase.

References

[1] Kreps, D. M., Milgrom, P., Roberts, J., and Wilson, R. Rational Cooperation in the Finitely Repeated Prisoners' Dilemma. Journal of Economic Theory, 27(2):245–252, 1982.

[2] Fudenberg, D. and Maskin, E. The Folk Theorem in Repeated Games with Discounting or with Incomplete Information. Econometrica, 54(3):533–554, 1986.

Questions for authors

Q1. What exactly is the intended meaning of the term "Memory Curse"?
Is the claim that longer memory simply reduces cooperation, or that it causes strategically worse behavior? This distinction matters because the current Memory Cursed categorization is defined by a decrease in cooperation rather than by lower reward or strategic suboptimality.

Q2. How should the reported reward results be interpreted relative to the cooperation-based definition of the curse?
The paper reports both discounted and realized rewards, and in some settings lower cooperation is accompanied by lower reward. However, the scripted Always-Defect control also shows that higher cooperation can reduce payoff substantially. Could the authors clarify whether the central claim is about cooperation itself or about failure to maximize long-run reward?

Q3. In the main self-play experiments, how do the authors distinguish an individual memory effect from an endogenous shift in the joint interaction regime?
Because agents with the same backbone interact repeatedly and share the same memory condition, increasing history length changes the behavior of all players simultaneously. Could the observed low-cooperation regime arise from reciprocal adaptation and retaliation rather than from one agent misusing memory?

Q4. Why are the scripted-opponent experiments not used to directly test the effect of memory length against a fixed opponent policy?
The paper already evaluates Always-Defect, Tit-for-Tat, and One-error Tit-for-Tat opponents, but these controls vary prompting rather than comparing short- and long-memory agents against the same fixed policy. Such a comparison seems important for separating opponent-dependent adaptation from a harmful memory effect.

Q5. What evidence would the authors consider sufficient to show that long-memory agents are overweighting stale negative events rather than rationally conditioning on opponent history?
For example, if an opponent defects early but later becomes persistently cooperative, would continued retaliation by the long-memory agent be considered the intended form of the "curse"? Clarifying this would help distinguish appropriate history-dependent adaptation from genuinely maladaptive persistence.

9 comments
AuthorsOpenReview ↗

Response to Reviewer 3Ura (1/5)

Thank you for your thoughtful game-theoretic perspective. Your comments highlight an important distinction between the collective behavior emerging when all agents adopt a shared memory configuration and the strategic performance of an individual agent against given opponents. Our primary experiments investigate the former, and we have added controlled fixed-opponent experiments to directly examine the latter.

W1(i) / Q1. What does the Memory Curse mean?

Our primary contribution is to characterize how the collective behavior of interacting LLM agents changes when the entire group adopts a longer memory window, rather than to evaluate the optimality of an individual agent's unilateral strategy.

Specifically, for each model and game, we compare two system-wide configurations:

  • (All) HL=2: Every agent has access to the two most recent interaction rounds.
  • (All) HL=80: Every agent has access to the 80 most recent interaction rounds.

The model backbone, group composition, and game rules remain fixed. What changes is the memory configuration of every agent in the interacting group, simultaneously.

We use Memory Curse to describe the resulting decline in collective cooperation under the expanded-memory configuration. Our central research question is therefore:

When all agents in a repeated multi-agent system gain access to substantially more interaction history, how does the cooperation sustained by that system change?

This is distinct from asking whether a single agent computes a best response to its opponents or whether a particular strategy profile constitutes a Nash equilibrium.

Our system-level perspective is important because memory is a design choice whose consequences emerge through interactions among agents. Even when agents respond individually to their observed histories, their collective behavior can change substantially when the memory available to the entire group expands. Understanding these emergent patterns is therefore valuable for designing cooperative multi-agent systems.

We will clarify this definition in the Introduction and Experimental Setup, and revise Figure 1's "Algorithmic Dysfunction" wording to describe the observed collective cooperation decline more precisely.

W1(ii) / Q2. Cooperation, realized payoffs, and welfare significance

The Memory Curse is defined by a decline in collective cooperation, but our payoff evaluations show that this decline is also strongly associated with lower realized returns.

The Appendix Tables 2–5 of our paper report discounted and undiscounted per-player rewards across all seven models and four games. Focusing on the Reasoning (CoT, labelled R in the tables) condition, we compare mean rewards at HL=2 and HL=80:

Table: Changes in mean per-player reward from HL=2 to HL=80 (CoT; Appendix Tables 2–5).

Experimental settingsDiscounted reward (Tables 2–3)Mean relative decreaseUndiscounted per-round reward (Tables 4–5)Mean relative decrease
All 28 model–game settings20 decrease, 6 unchanged, 2 increase (both only +0.03)13.7%21 decrease, 6 unchanged, 1 increase (only +0.03)17.1%
18 Memory Cursed settings17/18 decrease; sole increase only +0.0321.1%18/18 decrease26.4%

Mean relative decrease is the equally weighted average of each setting's relative reward reduction, including unchanged and increasing cases. Counts refer to directions of three-run means at the reported precision.

These results establish a strong correspondence between declining cooperation and lower realized payoffs. Among the 18 Memory Cursed settings, discounted reward decreases in 17/18 cases, with an average relative reduction of 21.1%, while undiscounted per-round reward decreases in all 18 cases, averaging 26.4%. Thus, the Memory Curse is typically accompanied by a substantial deterioration in the group's realized returns, extending its significance beyond cooperation rates alone.

(response to this question to be continued)

AuthorsOpenReview ↗

Response to Reviewer 3Ura (2/5)

(continued)

We deliberately select repeated social dilemmas in which mutual cooperation provides a clear, welfare-efficient, and mutually beneficial outcome, despite incentives for unilateral defection. In our payoff specifications, mutual cooperation maximizes total per-round welfare and offers every participant higher long-run returns than persistent mutual defection. This makes cooperation a meaningful benchmark for collective performance rather than an arbitrary behavioral target. This benchmark is particularly appropriate for our history-based multi-agent setting, where agents coordinate through observed actions and payoffs without directly negotiating or committing to a joint strategy. More elaborate history-dependent conventions, such as alternating roles, conditional action schedules, or coordinated responses to deviations, may require agents to establish shared expectations and maintain coordination across many rounds. Deviations can further complicate these conventions by triggering retaliation or uncertainty about subsequent behavior. Mutual cooperation, by contrast, provides a simple, symmetric, and readily identifiable coordination target that benefits all participants without requiring complex role assignments or alternating action schedules. Importantly, sustaining this outcome may itself require appropriate history-dependent strategies, while in the meantime, we do not assume that unconditional cooperation is always an individually optimal response.

Our central finding is that expanding memory for every agent can undermine this mutually beneficial collective outcome and, in the vast majority of our tested settings, reduce the group's average realized payoff. This demonstrates why the Memory Curse matters beyond cooperation rates: a shared memory-design choice can affect agents' ability to sustain welfare-efficient interaction patterns, with measurable consequences for their realized returns.

Complementing this system-level analysis, our new fixed-opponent experiments below in our response to W3/Q4 directly evaluate individual strategic performance. They identify conditions where expanding only the focal agent's memory reduces its own realized return relative to an implemented short-memory alternative under the same opponent policy.

AuthorsOpenReview ↗

Response to Reviewer 3Ura (3/5)

(continued)

W2 / Q3. Joint memory interventions versus unilateral strategic deviations

The key distinction is between the collective consequences of a joint memory intervention and the individual consequences of a unilateral strategic deviation.

In a best-response analysis, the other players' strategies are held fixed while a focal player considers alternative strategies to maximize its own payoff. Establishing a Nash equilibrium additionally requires verifying that no player can profitably deviate.

Our primary experiment asks a different question: we change every agent's memory configuration simultaneously, from all-HL=2 to all-HL=80, and examine the resulting collective interaction pattern.

Under this joint intervention, agents' responses to one another are naturally endogenous. An initial defection may trigger retaliation, which enters the histories observed by other agents and influences their subsequent decisions.

Crucially, this reciprocal adaptation is part of the collective behavioral effect we aim to study, rather than a confounder in our primary comparison. Our goal is precisely to understand the total system-level consequences of changing a shared memory design, including the feedback between agents that shapes their resulting cooperation.

This perspective also connects naturally to the Folk Theorem (Fudenberg and Maskin, 1986). Repeated games can support multiple history-dependent equilibrium outcomes under appropriate conditions. Our empirical question is which behavioral patterns actual LLM agents produce when their shared memory configuration changes. The existence of multiple possible equilibrium outcomes does not determine which interaction pattern a particular population of LLM agents will realize.

Our action-transition analysis identifies a corresponding behavioral signature. The probability of remaining non-cooperative following a non-cooperative action increases from 61.75% at HL=2 to 90.42% at HL=80, whereas the probability of transitioning from cooperation to non-cooperation remains approximately unchanged (11.16% versus 10.19%, Table 2).

This shows that expanded memory is associated primarily with greater persistence of non-cooperative behavior once it occurs, rather than more frequent initiation of defection. Such persistence helps characterize the distinct collective interaction patterns that emerge under different memory configurations.

Nevertheless, your individual-level perspective provides a valuable complementary test. We therefore extend our evaluation to settings where only the focal agent's memory length changes while the opponent follows a fixed, explicitly specified decision rule (detailed below in our response to W3(i)/Q4). This directly addresses the individual strategic-performance question raised in your review.

W3(i) / Q4. Direct memory comparisons against fixed opponent policies

Our primary contribution is to characterize how the collective behavior of interacting LLM agents changes when the entire group adopts a longer memory window, rather than to evaluate the optimality of an individual agent's unilateral strategy. As clarified in W1, our main experiment compares all-HL=2 with all-HL=80, where every player's memory changes together. The original scripted-opponent controls examined prompting effects on cooperation and payoff. Following your suggestion, we have added a complementary comparison: varying only the focal agent's memory while holding its opponent's policy rule fixed, and directly measuring individual return.

We evaluate Mistral-7B and Qwen3-8B against eight scripted policies in repeated PD, including Always-Defect, Tit-for-Tat, One-error Tit-for-Tat, noisy policies, and time-varying opponents. The table reports the CoT evaluation, where each response contains reasoning and an action. Each trajectory lasts 200 rounds, with 20 paired runs for Mistral and 11–16 for Qwen.

Within each comparison, the opponent follows the same decision rule at both history lengths. For reciprocal policies such as Tit-for-Tat, its realized actions can still respond to the focal agent, but its underlying decision rule remains fixed.

(response to this question to be continued)

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

This study conducted a large number of controlled experiments using seven large language models and four variations of the repeated social dilemma game, and identified a “memory curse” phenomenon. This phenomenon—in which providing LLM agents with a longer interaction history often weakens cooperation—is analyzed from multiple perspectives, including behavioral inertia, memory content, and reasoning patterns.

Critical strengths
  1. The topic of this paper is highly valuable; the memory window of agents is currently a critical yet under-researched area.
  2. The phenomenon identified in the paper is quite novel. The discovery of the “memory curse”—which contradicts the prevailing belief that “the longer the memory, the better the AI’s cooperation”—provides valuable insights for the field of multi-agent cooperative AI.
  3. The experiments described in the main text and supplementary materials are thoroughly conducted. They include testing numerous different LLM models, with each experimental condition repeated multiple times to account for randomness, and each experimental conclusion is accompanied by a detailed analysis and description.
Critical weaknesses
  1. The authors describe a counterintuitive “memory curse,” whereby expanding the available interaction history can undermine cooperation. However, related phenomena have already been documented. For example, [1] report that longer memory suppresses cooperation in Gemini but promotes it in Gemma. [2] explicitly investigate the effect of memory window size. The manuscript should clarify its contribution relative to these findings.
  2. The main text provides insufficient background on LLM agents in social dilemmas and repeated games. A more thorough introduction to this setting and its key concepts would help readers understand the motivation before proceeding to the experimental design and setup.
  3. The reported memory curse depends strongly on the model and task, with some combinations remaining immune. Moreover, token length appears to be a surface-level factor, whereas the content of the interaction history plays a more direct role in shaping cooperation. These results raise questions about whether the memory curse constitutes a robust, generalizable phenomenon or a collection of configuration-dependent and potentially stochastic effects.
  4. Relatedly, the conclusions appear specific to cooperation in multi-agent social dilemma games. Many other tasks require access to extensive historical information. The scope of the conclusions should therefore be explicitly restricted to cooperation in these settings, and their relevance to broader agent applications should be discussed more carefully.
  5. The appendix reports very high variance and substantial sensitivity to random seeds in some settings without chain-of-thought (CoT). In such cases, increased cooperation may reflect less consistent execution of retaliatory strategies or greater behavioral randomness, rather than improved reasoning or an acquired capacity for forgiveness. A higher cooperation rate alone is insufficient to distinguish these explanations.

[1] Hishiki et al., 2026, How memory can affect collective and cooperative behaviors in an LLM-Based Social Particle Swarm

[2] Fontana et al., 2025, Nicer than Humans: How Do Large Language Models Behave in the Prisoner’s Dilemma?

Questions for authors
  1. To what extent can findings about memory length in social dilemmas and repeated games inform memory design for agents performing practical tasks? What connects these experimental settings to broader agent applications?
  2. Does a similar phenomenon occur in a broader range of API-accessible, closed-source agent models?
  3. Why does the same backbone exhibit opposite memory effects under different evaluation configurations? For example, in the Prisoner’s Dilemma, Mistral’s cooperation rate decreases from 96.9% at HL = 2 to 85.3% at HL = 80 in the primary experiment, whereas it increases from 84.4% to 91.1% in the local base cohort.
7 comments
AuthorsOpenReview ↗

Response to Reviewer WJku (1/5)

Thank you for recognizing the importance of agent memory and the value of our empirical findings. Your comments have motivated us to clarify our contribution relative to prior work, strengthen the analysis of behavioral robustness, and connect our findings more explicitly to practical multi-agent systems. We have also added some experiments on behavioral stability, closed-source models, and evaluation protocols.

W1. Novelty and relation to prior work

Our contribution goes beyond observing that memory length affects cooperation: we systematically investigate how expanded interaction history changes cooperation in fixed groups of LLM agents, characterize the associated behavioral dynamics, and evaluate ways to improve cooperation while retaining longer histories.

The studies you cite provide valuable complementary perspectives. Hishiki et al. (2026) examine memory-dependent collective behavior in a spatial swarm, where agents interact with changing neighbors. Fontana et al. (2025) investigate memory-window effects in iterated Prisoner's Dilemma against externally specified opponent strategies. Our primary setting instead consists of fixed groups of repeatedly interacting LLM agents, where all agents adopt the same memory configuration and each player's decisions influence the histories and subsequent behavior of the same partners.

Our study contributes three further insights.

First, we identify where cooperation deteriorates. Our action-transition analysis shows that longer memory is associated primarily with greater persistence of non-cooperation after defection, rather than a comparable increase in the initiation of defection. This provides a behavioral characterization of the Memory Curse beyond aggregate cooperation rates.

Second, we isolate the importance of remembered interaction content. By modifying older records while keeping the number of visible rounds fixed, and separately using token-matched memory-representation controls, we demonstrate that cooperation depends on what agents remember and how that information is presented, not merely on the number of past rounds available.

Third, we investigate how to improve cooperation without discarding useful history. Under a common repeated-game protocol, our rationale-only fine-tuned model achieves higher mean cooperation than both the unadapted long-history model and the recent-history truncation baseline across all four games. This provides an alternative memory-design direction: improving how agents respond to accumulated history rather than simply reducing the available information.

Together, these analyses connect the observed phenomenon to behavioral dynamics, controlled memory interventions, and concrete mitigation strategies. We will make these distinctions from prior work more prominent in the Introduction and Related Work.

W2. Background on repeated social dilemmas

We will add an accessible introduction to repeated social dilemmas before Section 3, beginning with a concrete Prisoner's Dilemma example.

In our implementation, mutual cooperation gives each player 200 points, whereas unilateral defection gives 300 to the defector and −100 to the cooperating player. This illustrates the tension between an individual's immediate incentives and outcomes that benefit both players.

Repetition introduces an additional strategic consideration: an action taken today can influence how the same partners respond in future rounds. Cooperation may be sustained through reciprocity, while defection and retaliation can influence subsequent interactions. The history available to an agent therefore directly affects the information it uses to decide how to respond.

We will also expand Figure 1 to show the actual action–payoff records visible at HL=2 and HL=80, clarifying that our primary experiments vary the accessible history for every agent in the interacting group.

AuthorsOpenReview ↗

Response to Reviewer WJku (2/5)

(continued)

W3. Robustness and the role of history content

Our experiments establish a recurring pattern of memory-dependent cooperation decline across models and incentive settings, and further show that the content of remembered interactions plays an important role in shaping this behavior.

Robustness across configurations. Our primary evaluation covers seven models, four games, nine history lengths, and three independent 500-round runs per condition. We identify 18 Memory Cursed and 10 Memory Immune model–game settings. Importantly, the cooperation decline persists under changes to game incentives: in Mistral's Public Goods payoff sweep, the HL=2 to HL=80 decline remains 21.2–24.6 percentage points across all nine tested multipliers (Appendix Table 27).

Our action-transition analysis reveals a consistent behavioral signature. The probability of remaining non-cooperative after a non-cooperative action increases from 61.75% at HL=2 to 90.42% at HL=80, while the probability of transitioning from cooperation to non-cooperation remains approximately unchanged (11.16% vs. 10.19%, Table 2). This suggests that expanded memory affects cooperation primarily through greater persistence of non-cooperative behavior once it occurs.

History content at a fixed history length. We directly examine the role of remembered interactions through a controlled intervention in Traveler's Dilemma. Both conditions expose agents to 80 historical rounds, retaining the two most recent real interaction records while filling the remaining 78 slots with either IDEAL (mutual cooperation) or POLAR (polarized actions including defection) synthetic histories.

Table: Cooperation (%) under different older-history contents at HL=80, mean ± between-run SD (Appendix H and Appendix Table 15).

ModelIDEAL historyPOLAR history
GPT-OSS-20B41.77 ± 2.633.10 ± 0.36
Mistral-7B52.40 ± 1.7128.83 ± 2.10

These large differences demonstrate that the specific interactions contained in memory can substantially change cooperation even when the number of visible rounds remains fixed. This provides direct evidence that the effect of memory is not adequately described by history length alone.

Memory representation and token-matched controls. We further examine this distinction in Section 4.7 and Appendix L (Tables 22–23). Using 240 matched decision contexts across four games, we compare the original 80-round history (Raw80) with compact recent, summary, retrieval, and token-matched raw-history views.

In the four-game aggregate, Recent2 improves cooperation over Raw80 by 9.4 pp, while Retrieval improves it by 3.9 pp. The structured summary and retrieval views achieve results comparable to their token-matched compact raw controls.

We additionally evaluate whether these improvements carry into repeated interaction. Starting from nine matched Public Goods history prefixes, we continue to play for another 100 rounds under different memory representations.

Table: Cooperation improvements over Raw80 in PG continuation experiments (pp; paired 95% CIs; Appendix Table 23).

Memory representationGain over Raw80
Summary + Recent2+13.4 [10.3, 16.3]
Token-matched compact raw history+11.9 [9.9, 13.8]

Both compact representations substantially improve subsequent cooperation, with positive confidence intervals. The comparable performance of token-matched raw histories also shows that such improvements do not necessarily require a specialized summarization mechanism.

Together, these experiments establish that memory-dependent cooperation is shaped by both the information agents retain and how they use it. Our findings therefore motivate evaluating memory systems not only by context length, but also by the interaction histories they expose and the cooperative behavior they induce. We will emphasize this distinction in the revised discussion.

AuthorsOpenReview ↗

Response to Reviewer WJku (3/5)

(continued)

W4 / Q1. Implications for practical agent systems

Our findings are relevant to an increasingly important class of applications: AI agents representing different users or organizations that must coordinate despite partially aligned incentives.

This is already emerging in agent-mediated commerce. CarEdge offers a consumer-facing AI Buying Agent (https://cs2.caredge.com/live) and dealership-facing AI products (https://caredge.com/for-dealers), and Mark Miller Subaru (https://www.markmillersubarumidtown.com/blog/ai-agent-car-sale) has reported an agent-to-agent vehicle transaction involving a CarEdge buyer agent and a dealership-built agent. Luminance (https://www.luminance.com/press-releases/luminance-showcases-worlds-first-completely-ai-powered-contract-negotiation) has demonstrated AI agents negotiating a contract on behalf of opposing parties, while Pactum (https://pactum.com/campaigns-agents) develops agents that negotiate commercial terms with suppliers.

These applications create opportunities for cooperation even when agents represent parties with different objectives. For example, buyer and dealer agents may disagree on price, yet both benefit from accurate transaction information and efficient issue resolution. An agent might invest additional time verifying quoted fees, correcting errors, or resolving delivery problems. Such efforts incur individual costs while generating shared benefits, resembling the incentives in Public Goods. In recurring relationships, previous assistance, cooperation, and defection can also influence future reciprocity, as in repeated Prisoner's Dilemma.

Importantly, historical interactions and reputation are important to such decisions. For instance, CarEdge's Dealer Transparency Index (https://caredge.com/rating-methodology) uses historical dealer quotes to assess pricing transparency, illustrating how past behavior can inform future evaluations of counterparties. Similarly, CoopEval (Tewolde et al., 2026) studies how repeated interaction and reputation can sustain cooperation among LLM agents.

This is precisely where our findings offer an important insight: although historical information can be valuable for cooperation and reputation, simply giving agents more interaction history does not necessarily improve their cooperative behavior. Our experiments reveal that expanded memory can instead reinforce persistent non-cooperation. Thus, the effectiveness of agent memory depends not only on how much history is retained, but also on which interactions are remembered and how agents respond to them.

The practical implication is that agent memory should be evaluated not only by information retention, but also by the cooperative behavior and realized payoffs it induces. For incentive-sensitive multi-agent workflows, our findings motivate comparing recent-history, expanded-history, and adapted long-history policies under relevant partner behaviors and incentives. This provides a concrete direction for designing agent memory systems that support effective long-term coordination.

We will strengthen the Introduction and Discussion with these real-world applications, their connections to repeated social dilemmas, and the resulting implications for memory design in cooperative multi-agent systems.

Reference: Tewolde, Emanuel, et al. "Coopeval: Benchmarking cooperation-sustaining mechanisms and llm agents in social dilemmas." Proceedings of the 43rd International Conference on Machine Learning (2026). PMLR 306, 120222–120287

Reviewer_UjNoOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper finds an interesting pattern called memory curse, where expanding the history memory can reduce cooperation in repeated social dilemma games.

This paper first conducts extensive experiments with seven LLMs, four games, nine history lengths to show the existence of memory curse. The main experiments show that many model–game settings cooperate less when the visible history expands from 2 to 80 rounds, although some remain highly cooperative. To further investigate the mechanisms underlying this pattern, the authors conducted a broad range of analyses and experiments, including estimating defection-to-cooperation and cooperation-to-cooperation transition probabilities, analyzing the vocabulary in the LLM thinking process, experimenting with asymmetric memory, cleaning interaction histories, ablating explicit reasoning, and exploring alternative memory representations. They also trained an action-balanced, rationale-only Mistral adapter and reported improved cooperation in the public goods game, along with positive shifts in average outcomes across three games not included in training.

Critical strengths
  1. The "memory curse" pattern found in this paper is interesting. Exploring the reasons behind this pattern is a worthwhile work.
  2. The paper presents extensive experiments with clearly described experimental settings and high reproducibility. These experiments analyze the "memory curse" pattern from different perspectives, and the results yield some very interesting conclusions. Given the interesting research question and the extensive experimental evaluation, I lean toward a weak accept recommendation. However, as I do not have direct research experience in this area, I am not fully confident in my assessment.
Critical weaknesses
  1. This paper emphasizes the discovery of the "memory curse" pattern in its abstract and introduction, which is also its main claim. However, the results in Figure 2 show that not all experiments exhibited the memory curse pattern, a fact acknowledged in line 195, which is termed "memory immune." These results indicate that the memory curse is not a universal phenomenon.

  2. Both memory curse and memory immune defined in this paper are presented within a specific "model + game" context, not inherent properties of the model. In my view, memory curse should be a possible attribute of the model, not an attribute under the "model + game" context, because memory is relative to the model and independent of the game. For example, this attribute could be graded according to the severity of the memory curse exhibited by a model across games. Under such a scheme, GPT-OSS-20B and GPT-OSS-120B would fall into the highest level of memory curse, whereas Llama-3.3-70B and Qwen2.5-Coder-32B would fall into lower level of memory curse, according to the result in Figure 2. It may be more informative to investigate why models exhibit different degrees of the memory curse. For instance, whether these differences stem from model architecture, training data, or training procedures.

Questions for authors
  1. Have the authors considered the second point raised in "Critical Weaknesses"? This reflect an alternative perspective on the memory curse, but it doesn't necessarily mean that any one perspective is correct. I simply find this perspective more compelling. If the authors have considered this perspective and conducted relevant experiments, I would be very interested in learning about their findings. If not, there's also no need to add new experiments, as this is just a discussion.
  2. Minors: In line 375, this paper uses PP and CC to denote the defensive and cooperative dictionary matches, however, PP and CC have been alreadly used before. Therefore, it would be better to use other unused symbols.
3 comments
AuthorsOpenReview ↗

Response to Reviewer UjNo (1/2)

Thank you for recognizing the significance of the Memory Curse and the breadth of our experiments. We particularly appreciate your suggestion to characterize memory susceptibility at the model level. We believe this provides a valuable complementary perspective that further strengthens the paper.

W1. Memory Curse and Memory Immunity

A central finding of our study is a striking empirical dichotomy in how LLM cooperation responds to expanded interaction history. Across 28 model–game settings, we identify two behavioral regimes:

  • Memory Immune (10/28): Cooperation remains consistently high (mostly above 90%) across the tested history lengths from HL=2 to HL=80.
  • Memory Cursed (18/28): Cooperation declines as the visible history expands, although the intermediate trajectories need not be strictly monotonic.

Importantly, we observe no substantial cooperation improvement from expanding an already short history. Comparing HL=2 with HL=80, cooperation is unchanged or lower in 25/28 settings, while the remaining three exhibit only marginal increases of 0.2–0.8 percentage points from near-ceiling baselines.

This is our central counterintuitive finding: giving agents substantially more interaction history does not necessarily improve cooperation and can instead substantially undermine it. Rather than observing the intuitively expected improvement from longer memory, we find either sustained high cooperation or a deterioration in cooperation across the tested configurations.

Our contribution goes beyond identifying these two regimes to examining the behavioral dynamics underlying the cooperation decline. As shown in Table 2, the probability of remaining non-cooperative following a non-cooperative action increases from 61.75% at HL=2 to 90.42% at HL=80, whereas the probability of transitioning from cooperation to non-cooperation remains approximately unchanged.

This reveals an important behavioral signature: expanded memory is associated primarily with greater persistence of non-cooperation once it occurs, rather than more frequent initiation of defection. Our subsequent reasoning, memory-content, and intervention analyses further investigate these memory-dependent behaviors.

Together, these findings establish that more accessible memory is not automatically better for cooperation, and highlight the importance of studying how agents use their interaction histories. We will make the two observed regimes and this central finding more prominent in the abstract and introduction, and clarify their definitions in Figure 2.

AuthorsOpenReview ↗

Response to Reviewer UjNo (2/2)

(continued)

W2 / Q1. Model-Level Memory Susceptibility

Thank you for this insightful suggestion. Your model-level perspective offers another useful way to examine the Memory Curse. We therefore extend our original model-game analysis with a model-level susceptibility measure, allowing us to compare how strongly different models are affected by expanded memory across strategic environments.

Specifically, we define a model-level memory susceptibility score as the average cooperation decline from HL=2 to HL=80 across the four evaluated games, weighting each game equally. This provides a quantitative measure of how strongly a model's cooperation is affected by expanded memory across the tested strategic environments.

Table: Cooperation decline from HL=2 to HL=80 (percentage points), calculated from the Appendix Table 1. Positive values indicate lower cooperation with longer history.

ModelPDPGTDTGMean drop
GPT-OSS-120B0.071.480.285.459.25
GPT-OSS-20B71.538.339.325.343.60
Gemma-3-12B0.021.035.041.724.43
Qwen2.5-Coder-32B0.01.380.9−0.320.48
Llama-4-Scout-17B0.036.8−0.842.819.70
Mistral-7B11.624.68.63.011.95
Llama-3.3-70B0.0−0.20.043.910.93

The results support the model-level differences you identified. GPT-OSS-120B and GPT-OSS-20B exhibit the largest average cooperation declines, whereas models such as Llama-3.3-70B and Qwen2.5-Coder-32B show substantially lower average susceptibility.

Importantly, this analysis also reveals why model-level and game-level perspectives are complementary. For example, GPT-OSS-120B maintains 100% cooperation in PD at both history lengths, yet its cooperation declines by 71.4 pp in PG and 80.2 pp in TD. Similarly, Qwen2.5-Coder-32B remains highly cooperative in PD and PG but exhibits a substantial decline in TD.

These findings suggest that memory susceptibility is shaped by both model-specific characteristics and the strategic environment in which the model operates. The model-level score summarizes susceptibility across games, while the game-specific results reveal how the same model responds to different incentives and interaction structures.

We also see your suggestion as an important direction for future research. The proposed susceptibility score provides a quantitative foundation for systematically investigating which model properties contribute to robust cooperation under expanded memory. In particular, controlled comparisons across model architectures, pretraining data, and different post-training stages could help identify which factors make models more or less susceptible to the Memory Curse.

We will incorporate this model-level analysis into Section 4 alongside Figure 2 and expand the Discussion to highlight these future directions. Together, the two perspectives provide a more comprehensive framework for understanding and eventually improving the robustness of LLM cooperation under expanded memory.

Q2. Minor: Notation

Thank you for pointing this out. We will use distinct notation for the four dictionary counts: ndefn_{\mathrm{def}}, ncoopn_{\mathrm{coop}}, nfuturen_{\mathrm{future}}, and nhistoryn_{\mathrm{history}}, consistently throughout the paper.

Thank you again for your thoughtful feedback. We hope our new results and clarifications address your concerns, and would greatly appreciate it if you could consider revisiting your rating.

I appreciate the authors' response and the additional experiments which address my concerns and I agree with the point that memory susceptibility is shaped by both model-specific characteristics and the strategic environment. I believe this work is valuable. The authors' response strengthen the paper’s contribution and my confidence. I will raise my rating to 4.

Reviewer_HZ7iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how the length of an agent’s visible interaction history (“memory”) affects cooperation between LLM agents in repeated social dilemmas. The authors evaluate 7 LLMs on 4 games across 9 history lengths, with three 500-round runs per setting.

The key finding is that in 18 of 28 model–game settings, cooperation is lower with long history (HL=80) than with short history (HL=2). The authors call this the “memory curse.” In the remaining 10 settings, cooperation stays near ceiling (“memory immune”). To analyse the effect, the authors use action-transition statistics, lexical analysis of rationales, mixed short- and long-memory groups, memory sanitisation and a chain-of-thought ablation. To reduce it, they test a LoRA fine-tuned Mistral-7B, compact memory views and future-oriented prompting. The fine-tune and compact views raise cooperation, while prompting has no effect.

Critical strengths

The scale of the study is a strength: it is large for a well-targeted and well-specified problem. The decline in cooperation with longer history is clearly visible in a majority of settings (Figure 2). The conditional transition frequencies in Table 2 are particularly interesting because they show where the problem occurs. Longer history mainly makes agents fail to recover after defection: P(D|D) rises from 54% to 90%, while the rate of defecting after cooperating stays roughly flat. In my opinion, the fine-tuning result is the most interesting one. The adapter raises mean cooperation across history lengths in the training game (+25.1 pp in Public Goods), and the gains appear to transfer zero-shot to the other three games.

Critical weaknesses

The problem is well identified, but the paper lacks a coherent theme about how to address it. Fine-tuning, alternative memory representations and prompting are each evaluated under a different protocol and mostly on Mistral-7B, so the approaches cannot be compared. As a reader, I cannot tell whether fine-tuning is better than simply giving the agent a compact recent history. A direct comparison on a common setup would substantially strengthen the paper. Table 3 report no variance or significance testing. The transfer gains are small (+1.5, +5.0 and +9.0 pp) and based on three runs each. It is unclear whether the transfer effects are real.

Lexical analysis used to motivate the forward-looking fine-tuning is never tested: the forward-looking language of the fine-tuned model’s rationales is not measured. There is therefore no evidence that the adapter changed the reasoning as intended, or that such a change explains the cooperation gains.

Questions for authors

n/a

3 comments
AuthorsOpenReview ↗

Response to Reviewer HZ7i (1/2)

Response to Reviewer HZ7i

Thank you for recognizing the value of our action-transition analysis and fine-tuning results. Your comments highlight a central question: Can we improve cooperation while retaining longer interaction histories, rather than simply discarding older information? We address this through a common-protocol comparison of interventions, additional evaluation and training replications, and direct measurements of post-training reasoning.

1. Fine-tuning versus shorter history and future-oriented prompting

Fine-tuning can improve cooperation beyond simply providing a shorter history. In the single-call Public Goods (PG) evaluation, base cooperation declines from 41.18% at HL=2 to 18.13% at HL=80. Our fine-tuned model achieves 54.73% at HL=80, exceeding the short-history baseline by 13.56 percentage points (95% CI [10.05, 17.06]). An independent confirmation using 12 additional 500-round runs per condition yields a consistent gain of 12.01 pp [10.90, 13.11].

Importantly, our adapter is trained on action-balanced PG rationales without direct final-action supervision. Thus, the improvement is achieved while preserving the 80-round history window and without directly training the model to output cooperative actions. This suggests an alternative to memory truncation: adapting how agents respond to the history they retain.

Following your suggestion, we further evaluate fine-tuning, recent-history truncation, and future-oriented prompting under shared settings.

  • Base80: The original, non-fine-tuned model with access to the most recent 80 rounds of interaction history.
  • FT80: Our rationale-only fine-tuned model with access to the same 80-round history.
  • Recent2: The original model with access to only the two most recent rounds.
  • Future80: The original model with access to 80 rounds, augmented with a future-oriented prompting instruction.

All conditions follow the same evaluation protocol. We evaluate three independently trained adapters in separate blocks, each with four 500-round rollouts per condition and game. This comparison directly tests whether fine-tuning improves cooperation beyond simply shortening history or encouraging future-oriented reasoning through prompting.

Table: Cooperation gains of FT80 over the three corresponding baselines (percentage points; 95% confidence intervals across three training blocks).

GameFT80 − Base80FT80 − Recent2FT80 − Future80
PG+25.03 [17.70, 32.37]+30.94 [23.82, 38.06]+17.69 [10.16, 25.23]
PD+17.28 [15.30, 19.25]+30.03 [26.27, 33.79]+18.92 [14.30, 23.53]
TD+42.56 [24.91, 60.20]+15.68 [−2.34, 33.69]+49.36 [28.61, 70.11]
TG+15.95 [11.59, 20.31]+27.65 [24.76, 30.54]+13.76 [9.09, 18.43]

FT80 achieves the highest mean cooperation in all four games. These comparisons provide a coherent evaluation of the interventions and show that training agents to use longer histories can outperform simply restricting their accessible history. We will emphasize this design principle in the revised intervention discussion. Our existing fixed-context comparisons of summary and retrieval views (Table 4) provide complementary evidence about memory representation.

2. Variability and zero-shot transfer

To address the uncertainty in the original Table 3, we will report the fine-tuning (FT) gains with evaluation-run variability at HL=2, HL=80, and averaged equally across all six tested history lengths.

Table: FT-minus-base cooperation gains (pp), mean ± SD for the original adapter.

GameGain at HL=2Gain at HL=80Mean gain across HL
PG (training)+19.53 ± 1.84+34.07 ± 2.60+25.07 ± 0.69
PD+1.73 ± 1.31+3.37 ± 1.37+1.49 ± 1.13
TD−0.47 ± 4.87+10.30 ± 9.96+5.03 ± 2.44
TG+5.13 ± 1.98+20.23 ± 1.47+8.98 ± 0.71

The original Table 3 evaluates one adapter across multiple history lengths. The new common-protocol experiment additionally examines training reproducibility: all three independently trained adapters improve HL=80 cooperation over Base80 in all four games.

Since fine-tuning uses only PG rationales, the consistent gains in PD, TD, and TG under the shared protocol provide additional evidence of cross-game transfer. We will report the original evaluation variability and the new training-block results separately to make both sources of evidence explicit.

AuthorsOpenReview ↗

Response to Reviewer HZ7i (2/2)

(continued)

3. Does fine-tuning produce more forward-looking reasoning?

Yes. Our original paper already independently validates the forward-looking reasoning measure, and our new experiments directly confirm that fine-tuning increases future-oriented reasoning in the model's newly generated rationales.

First, our analysis goes beyond lexical keyword counting. In the submitted paper (Section 4.2 and Appendix D.3), we validate our lexical measure using a blinded DeepSeek-V4-Pro judge that independently evaluates 2,016 complete rationales for future orientation on a 1–10 scale. Rationales in the top decile of our lexical forward-looking ratio receive a mean judge score of 6.12/10, compared with 4.03/10 in the bottom decile (Δ = +2.09, permutation p < 0.0001). This positive association holds across all seven models, providing independent semantic validation of the lexical measure.

Moreover, at HL=80, Memory Immune settings exhibit a substantially higher mean forward-looking ratio than Memory Cursed settings (0.504 vs. 0.339; Figure 3). This connects expressed future-oriented reasoning with the cooperation patterns identified in our main experiments and motivates our rationale-only fine-tuning intervention.

Second, following your suggestion, we now directly evaluate rationales generated after fine-tuning. We collect base and fine-tuned rationales across 240 matched decision contexts spanning all four games, using three independently trained adapters. A separate blinded Qwen3 evaluator scores future orientation on a 0–3 scale without access to model identities or final actions. We also apply our original lexical dictionaries to the same generated rationales.

Table: Changes after fine-tuning (FT minus base; 95% CIs across three training blocks). Keyword frequencies are normalized per 1,000 rationale words.

GameΔ Future-orientation rating (0–3)Δ Future words / 1,000Δ Cooperative words / 1,000
PG+0.272 [0.044, 0.500]+4.04 [1.28, 6.79]+8.56 [1.81, 15.32]
PD+0.333 [0.035, 0.632]+2.89 [−3.41, 9.20]+13.64 [4.18, 23.11]
TD+0.333 [0.184, 0.483]+1.62 [−1.93, 5.17]+9.99 [4.84, 15.14]
TG+0.450 [0.160, 0.740]+2.99 [2.04, 3.94]+10.66 [5.62, 15.69]

The key finding is that blinded future-orientation ratings increase in all four games, with positive 95% confidence intervals throughout. Importantly, these changes arise from rationale-only fine-tuning without direct final-action supervision. Thus, the improvement is not merely an increase in cooperative action labels during training: the adapted model also generates measurably more future-oriented rationales when evaluated on held-out decision contexts.

Together, the original semantic validation and our new post-training evaluations establish a consistent connection between future-oriented reasoning and cooperative behavior. Our findings show that rationale-only adaptation can simultaneously improve expressed future-oriented reasoning and cooperation, offering a promising way to address the Memory Curse while preserving longer interaction histories. We will incorporate these post-training measurements into the revised paper, highlighting the reasoning changes as evidence consistent with the proposed adaptation mechanism rather than definitive proof of causal mediation.

Thank you again for your thoughtful feedback. We hope our new results and clarifications address your concerns, and would greatly appreciate it if you could consider revisiting your rating.

Thank you again for your thoughtful feedback and for recognizing the value of our action-transition analysis and fine-tuning results. We have now revised the manuscript PDF and incorporated the additional experiments conducted during rebuttal, directly addressing your main concerns.

In particular:

  • Direct intervention comparison: We now compare fine-tuning, history truncation, and future-oriented prompting under a common protocol with three independently trained adapters. FT80 achieves higher mean cooperation than Recent2 and Future80 across all four games (Appendix R.2).
  • Variability and generalization: We report evaluation variability and confidence intervals, along with additional independent training and evaluation replications. All three adapters improve long-history cooperation in the three games excluded from training, providing additional evidence of zero-shot transfer (Table 4; Appendix R).
  • Changes in reasoning: We directly evaluate rationales generated after fine-tuning using 240 matched contexts and a blinded evaluator. Future-orientation ratings increase across all four games, with positive 95% confidence intervals throughout (Appendix R.3).

We believe these additional results substantially strengthen the paper's contribution: beyond identifying a counterintuitive memory-dependent phenomenon, we demonstrate that rationale-only adaptation can improve cooperation while preserving long interaction histories, offering concrete insights for memory design and future incentive-aligned model training.


We hope the new evidence and substantial manuscript revisions address your concerns. We would greatly appreciate your reconsideration of the paper's overall rating. Thank you again for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.