acceptodds
Under review as a conference paper at ICLR 2027

FlowState: Execution State as Memory for Long-Horizon LLM Agents

Abstract

Long-horizon tasks require LLM agents to continually draw on information from earlier interactions. However, retaining the full history increases context costs, while compressing it risks losing details needed later, and the relevance of historical information often becomes apparent as the task progresses. To address these challenges, we propose , which treats execution state as memory that can be retained and revisited across requests, unifying current decision-making with the reuse of historical information. FlowState preserves semantically typed state nodes, their relations, and references to raw tool observations, separating persistent retention from on-demand access. Within a single execution loop, Incremental State Update () maintains the current state based on new inputs and feedback, while Progressive State Access () progressively reveals historical states and supporting evidence as needed during reasoning. Together, these mechanisms enable agents to reassess prior decisions in light of new information and guide subsequent actions. Compared with a full-context baseline using the same DeepSeek-V4-Flash model, FlowState improves the average success rate on MemoryArena and the average pass rate on -Bench by and percentage points, respectively, while reducing total token consumption by % and %. These results demonstrate the performance and efficiency advantages of FlowState on long-horizon tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM Agents, Long-Horizon Task, Agent Memory

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fknDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FlowState is a training-free framework that treats an LLM agent's execution state as its long-term memory. The runtime keeps a Persistent State Repository of semantically typed state nodes (user preferences, environment knowledge and attributes, agent judgements, artifacts), their typed relations, and references to raw tool observations, but only places a compact Historical State Index of identifiers and one-line summaries in the model context. At each ReAct step the frozen policy jointly emits a state delta, a historical access request, and an action: Incremental State Update (ISU) adds, updates, or removes current-request nodes without regenerating the whole state, while Progressive State Access (PSA) expands selected historical nodes and, by following their disclosed relations, drills down to supporting evidence on demand. Against a Full Context baseline with the same DeepSeek-V4-Flash model, FlowState raises average success rate on MemoryArena by 4.55 points and average pass rate on τ3-Bench by 13.95 points while cutting total tokens by 43.2% and 40.6%, beats ReasoningBank, BM25, Mem0, ZipAct, and ZipAct+BM25 on most MemoryArena metrics, and shows similar gains with GPT-5.6-terra and Qwen3.5-397B-A17B. Ablations indicate ISU is needed for task success and PSA for token efficiency, and per-subtask analysis shows FlowState retains accuracy at later subtask depths where baselines collapse, though absolute success rates on Shopping and Travel remain near zero for all methods and no variance across runs is reported.

Critical strengths
  • Clear and well-motivated design principle. The paper identifies a real gap between state-tracking methods that compress away details and memory systems that retrieve blindly. Its answer, keep everything persistently but expose only an index and let the agent expand on demand, is simple and addresses the stated failure modes directly. The hotel example in Figure 1 makes the problem concrete.
  • Gains in both accuracy and token cost at once. Most methods trade one for the other. FlowState improves over Full Context on eight of nine MemoryArena performance metrics and both τ3-Bench domains while cutting tokens by roughly 40%. The comparison is fair in the sense that both use the same frozen model and the same tasks, tools, and scoring.
  • Ablations support both components. Replacing ISU with rule-built boundary nodes drops success rate to zero in both tested environments, and exposing all history instead of PSA costs more tokens and hurts Travel. This is reasonable evidence that the two mechanisms play distinct roles.
Critical weaknesses
  • Prompt confound between FlowState and baselines. FlowState runs with a long, carefully engineered instruction block and a structured four-field output protocol. The baselines use the benchmarks' released prompts. Part of the gain may come from better instructions and forced structured reasoning rather than from the memory architecture. A prompt-matched control, for example Full Context with the same ISU instructions but no persistence, is missing.
  • Baselines are evaluated on one model only. The cross-model claim compares FlowState to Full Context on three models. All memory and state-tracking baselines are DeepSeek-V4-Flash only, so there is no evidence FlowState beats the field on GPT-5.6-terra or Qwen.
  • Ablations are replacements, not removals, and do not isolate design choices. The no-ISU variant builds a single node from the final response per query, which is a deliberately weak alternative, so the SR-to-zero result mostly shows that one summary node is insufficient. The no-PSA variant exposes all history and still gets the best Shopping PS. Nothing tests the five-type taxonomy, the relation types, raw observation references, or the one-step disclosure delay.
  • Near-zero absolute success on the two flagship environments. Shopping SR is 5.33% and Travel SR is 0.74% for the best method. The ranking rests on progress scores whose definitions differ across environments, and the paper does not discuss why all methods fail so badly.
Questions for authors
  1. Are all reported results single runs? Can you provide results over at least three seeds with standard deviations, especially for Travel SR (one vs two tasks out of 270), Physics SR (one task out of 20), and Math SR (three tasks out of 40)?
  2. Could you add a prompt-matched control: Full Context with the same instruction block and four-field output protocol as FlowState, but without the repository, index, or PSA? This would separate the effect of the memory architecture from the effect of the engineered prompt.
8 comments
AuthorsOpenReview ↗

Response to fknD (1/4): Prompt-matched controls and cross-model baselines

Thank you for your careful review and for recognizing FlowState's design motivation, performance and cost benefits, and the roles of its two core mechanisms. We appreciate your suggestions on prompt-matched controls, cross-model comparisons, and ablations. We have carefully considered these concerns and provide additional results and explanations below.

W1 and Q2: Separating prompt effects from the contribution of the memory architecture

Prompt confound between FlowState and baselines.

Could you add a prompt-matched control: Full Context with the same instruction block and four-field output protocol as FlowState, but without the repository, index, or PSA? This would separate the effect of the memory architecture from the effect of the engineered prompt.

Thank you for this suggestion. To separate the contributions of the prompt and the memory architecture, we added a Full Context baseline on Math and Physics with a prompt matched to FlowState: Prompt-matched Full Context.

This baseline uses FlowState's system prompt and four-field output protocol from Appendix B.3 and retains the full conversation history. It records state_delta and state_focus only as ordinary messages, without executing the corresponding state operations or implementing the Persistent State Repository, Historical State Index, or PSA mechanism.

All methods in this comparison use DeepSeek-V4-Flash. The table reports mean ± standard deviation, with SR and PS expressed as percentages and token consumption in millions.

MethodMath SRMath PSMath tokens (M)Physics SRPhysics PSPhysics tokens (M)
Full Context25.83 ± 3.8249.13 ± 1.5137.36 ± 2.9153.33 ± 2.8969.40 ± 4.123.73 ± 0.12
Prompt-matched Full Context31.67 ± 5.7752.37 ± 4.2746.39 ± 3.7150.00 ± 10.0064.05 ± 2.895.99 ± 0.45
FlowState42.50 ± 5.0055.77 ± 0.2512.95 ± 1.7361.67 ± 2.8974.88 ± 2.101.98 ± 0.18

FlowState retains higher task performance and lower token consumption after matching the prompt and four-field output protocol. Compared with Prompt-matched Full Context, FlowState improves mean SR by 10.83 and 11.67 percentage points and PS by 3.40 and 10.83 percentage points on Math and Physics, respectively, while reducing token consumption by 72.1% and 66.9%. This comparison supports the contribution of the memory architecture beyond the prompt and structured output: by continuously maintaining reusable states and accessing historical information on demand, FlowState improves both task performance and context efficiency.

W2: Comparing memory and state-tracking baselines across models

Baselines are evaluated on one model only. The cross-model claim compares FlowState to Full Context on three models. All memory and state-tracking baselines are DeepSeek-V4-Flash only, so there is no evidence FlowState beats the field on GPT-5.6-terra or Qwen.

Thank you for this suggestion. To test whether FlowState's advantages over different types of baselines extend beyond a single model, we added a full baseline comparison on Math and Physics using Qwen3.5-397B-A17B, covering Full Context, ReasoningBank, BM25, Mem0, ZipAct, and ZipAct+BM25. The table reports mean ± standard deviation, with SR and PS expressed as percentages and token consumption in millions.

Qwen3.5-397B-A17B

MethodMath SR ↑Math PS ↑Math #Tok. ↓Physics SR ↑Physics PS ↑Physics #Tok. ↓
Full Context34.17 ± 3.8245.61 ± 9.0520.05M ± 0.40M48.33 ± 2.8960.51 ± 5.192.99M ± 0.86M
ReasoningBank30.83 ± 5.2046.63 ± 0.7210.99M ± 0.07M23.33 ± 2.8946.46 ± 2.892.67M ± 0.01M
BM2537.50 ± 5.0049.41 ± 2.1910.75M ± 0.05M45.00 ± 15.0060.18 ± 5.532.38M ± 0.08M
Mem027.50 ± 4.3343.15 ± 0.857.91M ± 0.05M16.67 ± 2.8944.26 ± 2.641.90M ± 0.08M
ZipAct13.33 ± 3.8221.79 ± 2.0215.76M ± 0.38M30.00 ± 5.0041.43 ± 4.613.23M ± 0.03M
ZipAct+BM2528.33 ± 2.8940.23 ± 2.0114.22M ± 0.09M36.67 ± 2.8953.30 ± 4.103.02M ± 0.12M
FlowState (Ours)38.33 ± 1.4450.46 ± 3.1314.18M ± 0.40M51.67 ± 7.6461.43 ± 5.282.38M ± 0.35M

As shown above, FlowState achieves the highest mean SR and PS in both domains with Qwen3.5-397B-A17B, while using fewer tokens than Full Context. Together with the original DeepSeek-V4-Flash experiments, these results show that FlowState's performance advantages over multiple memory and state-tracking baselines extend to another model, providing further empirical evidence of its applicability across models.

AuthorsOpenReview ↗

Response to fknD (2/4): Controlled ablations of individual design choices

W3 (Part 1): Individual contributions of the design choices

Ablations are replacements, not removals, and do not isolate design choices. The no-ISU variant builds a single node from the final response per query, which is a deliberately weak alternative, so the SR-to-zero result mostly shows that one summary node is insufficient. The no-PSA variant exposes all history and still gets the best Shopping PS. Nothing tests the five-type taxonomy, the relation types, raw observation references, or the one-step disclosure delay.

Thank you for this suggestion. We acknowledge that the original ablations primarily compared components and were insufficient to isolate the contribution of each design choice. To address this concern, we added five controlled variants on Math and Physics. Each starts from the full FlowState framework and changes only one mechanism:

  • No evidence access: The agent can still read historical states but cannot revisit the raw tool outputs supporting them. State content, relations, and current observations are retained.
  • No relation navigation: When a historical state is disclosed, links to its related states are hidden. The agent can still select and disclose historical states through the Historical State Index and independently access raw evidence.
  • No node types: We collapse the five state types into a single generic type, keeping all other design choices unchanged.
  • No relation types: We replace the relation labels between states with a single generic type, keeping all other design choices unchanged.
  • Full-state rewriting: We keep the update timing unchanged but require the agent to regenerate all modifiable states for the current request, rather than returning only additions, updates, and removals.

SR and PS are reported as percentages, and token consumption is reported in millions. Results are presented as mean ± standard deviation.

VariantMath SRMath PSMath tokens (M)Physics SRPhysics PSPhysics tokens (M)
Full FlowState42.50 ± 5.0055.77 ± 0.2512.95 ± 1.7361.67 ± 2.8974.88 ± 2.101.98 ± 0.18
No evidence access39.17 ± 3.8253.68 ± 1.7014.01 ± 0.4353.33 ± 12.5869.38 ± 3.852.10 ± 0.04
No relation navigation40.83 ± 1.4453.44 ± 1.3813.65 ± 0.1755.00 ± 10.0071.06 ± 5.402.05 ± 0.03
No node types41.67 ± 1.4453.97 ± 2.0013.70 ± 0.0553.33 ± 2.8968.87 ± 3.662.03 ± 0.07
No relation types40.83 ± 1.4455.93 ± 0.6713.87 ± 0.2850.00 ± 5.0069.42 ± 2.172.13 ± 0.06
Full-state rewriting31.67 ± 1.4454.58 ± 1.5413.72 ± 0.1958.33 ± 2.8972.80 ± 2.681.91 ± 0.13

As shown above, all five variants have lower mean SR than full FlowState in both domains. With access to historical states retained, the first four comparisons support the contributions of node types, relation semantics, relation navigation, and access to raw evidence. Full-state rewriting reduces mean SR despite preserving the update timing and the scope of modifiable states, further supporting the value of incremental maintenance. These experiments connect the overall gains to specific design choices, going beyond component-level replacements.

AuthorsOpenReview ↗

Response to fknD (3/4): Disclosure timing and benchmark difficulty

W3 (Part 2): Historical access and disclosure timing

Nothing tests the five-type taxonomy, the relation types, raw observation references, or the one-step disclosure delay.

Regarding the one-step disclosure delay, we clarify that historical information becomes visible in the next model invocation after an access request, but the agent can request retrieval alone and wait for the information before acting. When the current action does not depend on the information to be retrieved, the agent can also request historical information and execute an action in the same step.

To assess whether mandatory waiting is beneficial, we compare two scheduling settings:

  • Joint scheduling: The agent chooses whether to retrieve historical information alone or to request retrieval and execute an action in the same step.
  • Sequential scheduling: Every retrieval request pauses environment actions until the historical information is returned, after which the agent makes its next decision.

In both settings, historical information is disclosed in the next model invocation. The only difference is whether retrieval and environment actions must be executed sequentially.

SR and PS are reported as percentages, and token consumption is reported in millions. Results are presented as mean ± standard deviation.

SchedulingMath SRMath PSMath tokens (M)Physics SRPhysics PSPhysics tokens (M)
Joint scheduling42.50 ± 5.0055.77 ± 0.2512.95 ± 1.7361.67 ± 2.8974.88 ± 2.101.98 ± 0.18
Sequential scheduling40.00 ± 2.5055.93 ± 0.1913.85 ± 0.3160.00 ± 8.6674.12 ± 3.332.05 ± 0.04

The two settings yield similar mean task performance, while joint scheduling uses slightly fewer tokens. In qualitative trajectory analysis, we observe that the agent retrieves first and acts afterward when it needs historical information to decide on an action. When the available information already supports the current action, it can act and request retrieval in the same step. These results support allowing the agent to decide when to wait, without requiring sequential execution for every historical access request.

W4: Absolute success rates and the interpretation of progress metrics

Near-zero absolute success on the two flagship environments. Shopping SR is 5.33% and Travel SR is 0.74% for the best method. The ranking rests on progress scores whose definitions differ across environments, and the paper does not discuss why all methods fail so badly.

Thank you for highlighting this issue. The key difficulty in Shopping and Travel is integrating local decisions across sessions into a complete solution that satisfies all interdependent constraints. In Shopping, subsequent purchases must satisfy compatibility and preference constraints associated with earlier items. In Travel, each subtask contains 30 itinerary entries and involves dependencies across travelers and activities. Correctly completing some purchases or satisfying some itinerary constraints therefore does not guarantee that the final product bundle or overall itinerary passes evaluation. The original MemoryArena paper reports the same phenomenon: all methods in its Table 3 have zero Travel SR, and the listed external-memory and RAG methods also have zero Shopping SR [1].

Methods are compared under consistent evaluation rules within each environment. We fully adopt the evaluation rules and SR, PS, and sPS definitions of the MemoryArena implementation. All methods use the same tasks and completion criteria within each environment, so each comparison in the main results table rests on a consistent evaluation basis.

FlowState's results illustrate the gap between task progress and complete success. Higher progress scores reflect gains in completing individual subtasks or satisfying constraints, whereas low SR indicates that producing globally consistent product bundles and itineraries remains a major challenge. We therefore report both overall success rates and progress metrics to capture performance at the levels of complete task success and fine-grained progress.

[1] Zexue He et al. MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks. 2026.

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

The paper propose a new framework that treats execution state as memory that can be retained and revisited across requests, unifying current decision-making with the reuse of historical information. The framework maintains semantically typed state nodes and references to raw tool observations. During the inference, the ISU maintain the current state, and PSA reveals history state to support the reasoning. The results shows that the method bring improvement on different tasks.

Critical strengths
  • The research question is well-motivated. The long-horizon context management is an important task to investigate.
  • The method is intuitive and the experiments includes many different types of baselines.
Critical weaknesses
  • A more comprehensive analysis on the memory correctness will make the method's advantage more clear.
  • The scalability is another concern. As the history context getting longer and longer during the inference, the historical state index will be hard to maintain. It's unclear what is the cost of the method under long-horizon context with over 1000 records, which is the minimum number for today's long horizon agents.
  • Only Deepseek-v4-flash is evaluated in both benchmarks. From the current results, it seems memory arena is a relatively easy benchmark compared to another. It's remain unclear that whether the method is general enough across different model in more challenging benchmark.
Questions for authors
  • Why the travel planning success rate is relatively low? It seems like the method is far from the real-world application with this performance.
  • It seems like the token usage reduction is more impressive than the success rate. The method design seems didn't intentionally reduce the token usage, where this gain comes from?
6 comments
AuthorsOpenReview ↗

Response to kmFT (1/3): Memory quality and correctness

Thank you for your expert feedback and insightful comments. We have carefully considered them and provide additional results and explanations below.

W1: Memory quality and correctness analysis

A more comprehensive analysis on the memory correctness will make the method's advantage more clear.

Thank you for this valuable suggestion. We added a memory-quality evaluation covering all Math and Physics tasks, comparing FlowState, ReasoningBank, and ZipAct. We assess both the maintained memory and the information actually visible at decision time using three metrics:

  • Factual faithfulness: Whether the information in memory is supported by the original interactions.
  • Contextual completeness: Whether the retained information preserves the entities it concerns, its conditions, and its scope of applicability.
  • Recall sufficiency: Whether the visible memory contains all known historical information needed for the current decision.

The first two metrics evaluate the quality of the maintained memory; the third evaluates how well the memory visible at decision time covers the required historical information. We use DeepSeek-V4-Flash as the judge and hide method names during scoring. All metrics use a 100-point scale, with higher scores indicating better performance. For each judging pass, we first compute the mean score for each method. We then report the mean ± sample standard deviation of these means across three repeated judging passes.

EnvironmentMethodFactual faithfulness ↑Contextual completeness ↑Recall sufficiency ↑
MathReasoningBank90.65 ± 0.6889.89 ± 0.3167.26 ± 1.50
MathZipAct93.13 ± 0.5392.94 ± 0.7655.75 ± 0.49
MathFlowState94.05 ± 1.0494.14 ± 0.4279.23 ± 0.99
PhysicsReasoningBank87.62 ± 0.6487.06 ± 0.7562.85 ± 0.61
PhysicsZipAct90.69 ± 1.4889.93 ± 1.1249.86 ± 0.43
PhysicsFlowState96.99 ± 0.7296.35 ± 0.7985.09 ± 0.46

As shown above, FlowState achieves the highest mean score on all three metrics in both domains. Its largest advantage is in recall sufficiency: it improves over the strongest baseline in each domain by 11.97 points on Math and 22.24 points on Physics. These results show that FlowState provides more complete historical information for subsequent decisions while maintaining high factual faithfulness and contextual completeness, consistent with the goals of its state maintenance and progressive access mechanisms.

AuthorsOpenReview ↗

Response to kmFT (2/3): Long-history scalability and sources of token savings

W2: Scalability and index-maintenance costs with long histories

The scalability is another concern. As the history context getting longer and longer during the inference, the historical state index will be hard to maintain. It's unclear what is the cost of the method under long-horizon context with over 1000 records, which is the minimum number for today's long horizon agents.

Thank you for this suggestion. To directly evaluate FlowState's scalability beyond 1,000 historical records, we conducted long-horizon experiments on the Vending and Freelance tasks in EcoGym [1]. From actual execution traces, we measured history size, index size, and agent-side state-maintenance time, alongside task performance and cumulative token consumption.

Task performance with long histories. For Vending, we report the mean over three complete 365-day runs. For Freelance, we compare cumulative performance upon reaching day 255, when FlowState and Full Context have accumulated 2,806 and 2,399 business-history records, respectively. The results are as follows:

EnvironmentMethodNet worth / Balance ↑Business-history recordsTokens (M) ↓
VendingFlowState1,506.881,178.6724.34
VendingFull Context756.56811.0027.58
FreelanceFlowState1,490.572,806226.28
FreelanceFull Context252.112,399278.83

Business history is counted in business messages. Token consumption is cumulative over each trajectory, and all Vending metrics are averaged over three runs. Both Freelance methods are evaluated upon reaching day 255.

Index and state-maintenance overhead. During the Freelance run, as the business history grows from 500 to 2,801 records, FlowState's historical index grows from 435 to 2,683 entries, and its serialized size increases from 142 KB to approximately 1.00 MB. The mean agent-side maintenance times, grouped by history length, are shown below.

Business-history lengthMean agent-side maintenance (ms / environment step)
≤1,0001.54
1,001–2,0003.74
2,001–2,8067.18

Agent-side maintenance includes context construction, state updates, and tool-observation processing, but excludes model requests, logging to disk, and recovery and reconstruction. Although maintenance time increases with history length, it remains below 8 ms per environment step on average within the tested range of approximately 2,800 records.

These results show that FlowState supports both task performance and efficient context management with long histories. By locating historical states through a summary index and disclosing their full content on demand, FlowState supports long-horizon decision-making with over 2,800 historical records at millisecond-scale agent-side maintenance costs, while achieving higher returns and lower token consumption in both environments.

References

[1] Xavier Hu et al. EcoGym: Evaluating LLMs for Long-Horizon Plan-and-Execute in Interactive Economies. arXiv:2602.09514, 2026.

Q2: Where do the token savings come from?

It seems like the token usage reduction is more impressive than the success rate. The method design seems didn't intentionally reduce the token usage, where this gain comes from?

Thank you for this question. The main source of token savings is that FlowState compresses lengthy agent interaction histories into structured states that can be progressively disclosed. Subsequent requests can access historical information on demand through a compact index, without repeatedly including the complete interaction history in the context.

Specifically, ISU continuously maintains reusable states during execution and stores them in the Persistent State Repository. PSA progressively discloses relevant states and raw evidence according to the needs of subsequent requests.

https://anonymous.4open.science/api/repo/FlowState-F34D/file/context_growth_300dpi.png

The horizontal axis shows relative task progress, and the vertical axis shows the number of context tokens per LLM call on a logarithmic scale.

As shown in the figure, FlowState's context grows more slowly than Full Context, with the gap becoming more pronounced toward the end of execution. The median input lengths at task completion are shown below:

EnvironmentFull ContextFlowStateReduction
Math61.28K10.94K82.1%
Physics16.53K5.40K67.3%

Reducing repeated processing of the full history is therefore an important source of token savings. State reuse through ISU and on-demand access through PSA allow historical information to continue supporting reasoning while reducing the cost of repeatedly reading that information in subsequent calls.

AuthorsOpenReview ↗

Response to kmFT (3/3): Cross-model generalization and Travel performance

W3: Cross-model evaluation on more challenging interactive tasks

Only Deepseek-v4-flash is evaluated in both benchmarks. From the current results, it seems memory arena is a relatively easy benchmark compared to another. It's remain unclear that whether the method is general enough across different model in more challenging benchmark.

Thank you for this suggestion. To address cross-model generalization on a more challenging benchmark, we added evaluations of Qwen3.5-397B-A17B on the Airline and Retail domains of τ³-Bench, comparing FlowState with Full Context using the same model. Metrics follow the main paper, and results are reported as mean ± standard deviation, with performance metrics expressed as percentages.

DomainMethodPR↑DB↑Act.↑#Tok. ↓
AirlineFull Context86.00 ± 3.4689.78 ± 2.1388.28 ± 5.7216.60M ± 0.13M
AirlineFlowState89.33 ± 8.0891.99 ± 8.7092.23 ± 7.3313.75M ± 1.71M
RetailFull Context85.09 ± 2.3287.13 ± 2.0388.00 ± 0.4824.96M ± 0.10M
RetailFlowState93.27 ± 1.3494.15 ± 1.3489.82 ± 0.4826.51M ± 1.66M

As shown above, with Qwen3.5-397B-A17B, FlowState outperforms Full Context in mean PR, DB, and Act. in both Airline and Retail, consistent with the overall improvement trend reported for DeepSeek-V4-Flash in the main paper. These results show that FlowState's gains are not limited to a single foundation model, providing further empirical evidence of its ability to generalize across models on complex interactive tasks.

Q1: Why is the Travel success rate so low?

Why the travel planning success rate is relatively low? It seems like the method is far from the real-world application with this performance.

Thank you for raising this question. Travel is the most challenging environment in MemoryArena. It requires coordinating dependencies across travelers and activities as new travelers join over successive rounds, with each subtask requiring an itinerary containing 30 slots. All methods in the original benchmark's main results table have a Travel SR of 0% [1].

sPS measures the satisfaction of individual constraints, whereas SR requires the final itinerary to satisfy every traveler's requirements. Thus, even a small number of constraint violations can cause the entire task to fail. FlowState achieves an sPS of 91.56 ± 0.34%, but an SR of only 0.86 ± 0.21%. This indicates that it can satisfy most constraints but still struggles to ensure a fully valid overall plan. Although absolute task performance remains limited, FlowState's improvements in task performance and context efficiency under this demanding setting provide positive empirical evidence of its potential to support complex long-horizon tasks.

References

[1] Zexue He et al. MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks. 2026.

Reviewer_Qnb9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FlowState, a framework that unifies execution-state maintenance and persistent memory for long-horizon LLM agents. Incremental State Update (ISU) maintains structured states during the current request, while Progressive State Access (PSA) selectively exposes historical states and their supporting evidence. The framework separates persistent storage from context visibility and retains links between state representations and original tool observations.

Critical strengths
  1. Coherent representation and access design.
  2. Meaningful overall empirical improvements.
  3. Relevant comparisons and transparent reporting.
Critical weaknesses
  1. The reliability of the success-rate improvements is insufficiently established. Some improvements correspond to very few additional successful tasks. Based on the reported task counts and rounded success rates, Travel improves from approximately 1/270 to 2/270 successes, Physics from 11/20 to 12/20, and Math from 12/40 to 15/40. The manuscript does not provide repeated-run distributions or task-paired confidence intervals that would establish the stability of these differences.
  2. The ablations do not sufficiently connect the overall gains to the distinctive design choices. Removing ISU replaces rich, incrementally maintained states with final-answer nodes created after a request, changing update timing, information granularity, and content simultaneously. Removing PSA exposes all historical states, changing the amount of information in the context. These experiments evaluate useful component-level alternatives, but they do not isolate the benefits of typed states, relations, evidence links, or progressive access.
  3. The implications of historical-index growth remain unclear. The historical index includes summaries from all preceding requests, and disclosed historical states accumulate in the active state. Historical nodes cannot be removed through the same operations available for current-request nodes. Consequently, selective access alone does not establish that the visible context remains compact as the history grows.
Questions for authors
  1. How many runs underlie Tables 1–3, and can you provide task-paired outcomes and uncertainty estimates?
  2. Does PSA retain its advantage over single-pass retrieval when state content and input budgets are matched?
  3. How do index size and disclosed historical content grow, and what happens when they exceed the available context capacity?
8 comments
AuthorsOpenReview ↗

Response to Qnb9 (1/4): Repeated runs and task-paired analysis

Thank you for taking the time to review our paper and for recognizing FlowState's design and empirical improvements. We appreciate your expert feedback and have carefully considered your suggestions. Below, we provide additional experiments and clarifications.

W1 and Q1: Repeated runs and reliability of the results

The reliability of the success-rate improvements is insufficiently established.

How many runs underlie Tables 1–3, and can you provide task-paired outcomes and uncertainty estimates?

Thank you for this suggestion. The original manuscript reported a single run with seed 42. We have added runs with seeds 43 and 44 and report the mean ± standard deviation across the three seeds below. The evaluation settings and metric definitions are unchanged from the main paper.

Table 1: MemoryArena.

Model / MethodShopping SR ↑Shopping PS ↑Shopping #Tok. ↓Travel SR ↑Travel PS ↑Travel sPS ↑Travel #Tok. ↓
DeepSeek-V4-Flash · Full Context0.22 ± 0.3820.00 ± 1.0250.06M ± 3.47M0.37 ± 0.373.78 ± 0.6488.82 ± 0.47232.29M ± 0.68M
GPT-5.6-terra · Full Context1.33 ± 1.3337.22 ± 3.13295.31M ± 16.16M————
Qwen3.5-397B-A17B · Full Context0.00 ± 0.0018.63 ± 0.85140.24M ± 5.97M0.62 ± 0.215.94 ± 0.9944.26 ± 15.60249.81M ± 23.06M
ReasoningBank0.44 ± 0.7735.44 ± 3.79115.07M ± 19.40M0.00 ± 0.005.40 ± 1.7554.70 ± 5.7443.99M ± 4.64M
BM252.00 ± 1.3341.56 ± 16.8580.32M ± 41.19M0.00 ± 0.007.77 ± 2.3067.95 ± 2.4960.19M ± 8.35M
Mem00.22 ± 0.3832.30 ± 5.65111.39M ± 45.08M0.00 ± 0.005.53 ± 1.8743.62 ± 2.4961.35M ± 39.72M
ZipAct0.00 ± 0.0012.37 ± 4.6939.10M ± 4.45M0.00 ± 0.005.28 ± 1.2921.51 ± 3.37199.43M ± 2.96M
ZipAct+BM250.00 ± 0.0010.93 ± 4.5453.68M ± 10.60M0.00 ± 0.005.32 ± 0.9120.72 ± 0.86203.49M ± 24.36M
FlowState (Ours)3.78 ± 1.3942.44 ± 0.9749.74M ± 6.43M0.86 ± 0.2110.78 ± 2.1891.56 ± 0.34128.86M ± 2.72M
Δ vs. V4-Flash Full Context+3.56+22.44−0.64%+0.49+7.00+2.74−44.53%
Model / MethodMath SR ↑Math PS ↑Math #Tok. ↓Physics SR ↑Physics PS ↑Physics #Tok. ↓
DeepSeek-V4-Flash · Full Context25.83 ± 3.8249.13 ± 1.5137.36M ± 2.91M53.33 ± 2.8969.40 ± 4.123.73M ± 0.12M
GPT-5.6-terra · Full Context39.17 ± 3.8253.56 ± 0.7814.26M ± 1.49M53.33 ± 2.8967.81 ± 4.622.54M ± 0.18M
Qwen3.5-397B-A17B · Full Context34.17 ± 3.8245.61 ± 9.0520.05M ± 0.40M48.33 ± 2.8960.51 ± 5.192.99M ± 0.86M
ReasoningBank20.83 ± 10.4135.60 ± 8.724.69M ± 2.78M26.67 ± 24.6642.21 ± 19.151.15M ± 0.64M
BM2526.67 ± 5.2041.40 ± 7.346.48M ± 3.10M41.67 ± 5.7751.54 ± 7.871.31M ± 0.06M
Mem019.17 ± 5.2034.25 ± 6.914.58M ± 1.76M15.00 ± 18.0336.51 ± 14.780.95M ± 0.33M
ZipAct29.17 ± 1.4449.04 ± 1.5914.87M ± 2.24M48.33 ± 7.6466.77 ± 3.922.98M ± 0.32M
ZipAct+BM2528.33 ± 2.8950.05 ± 2.6314.54M ± 4.07M41.67 ± 2.8962.13 ± 2.022.85M ± 0.56M
FlowState (Ours)42.50 ± 5.0055.77 ± 0.2512.95M ± 1.73M61.67 ± 2.8974.88 ± 2.101.98M ± 0.18M
Δ vs. V4-Flash Full Context+16.67+6.64−65.34%+8.33+5.48−46.92%

Δ denotes the change relative to DeepSeek-V4-Flash Full Context, expressed in percentage points for performance metrics and relative percentages for token consumption.


Table 2: τ³-Bench.

DomainMethodPR ↑DB ↑Act. ↑#Tok. ↓
AirlineFull Context62.00 ± 4.0067.69 ± 2.9779.24 ± 2.8319.60M ± 1.93M
AirlineFlowState (Ours)85.33 ± 7.0286.53 ± 7.4889.58 ± 5.3513.27M ± 4.49M
RetailFull Context81.87 ± 7.1486.16 ± 4.7491.24 ± 0.8224.78M ± 1.09M
RetailFlowState (Ours)90.35 ± 7.6592.70 ± 5.9589.35 ± 3.3521.33M ± 2.67M

Task-paired analysis. We retain each task's paired outcomes for both methods across all three seeds and compute 95% paired bootstrap percentile intervals using deterministic convolution. These intervals reflect task-level uncertainty conditional on the three runs. "FS-only / FC-only successes" count task–run instances in which only FlowState or only Full Context succeeds, respectively, summed across the three runs.

Domain (number of tasks)FS-only / FC-only successesSR / PR difference (pp)95% paired interval (pp)
Shopping (150)17 / 1+3.56[+1.33, +6.00]
Travel (270)7 / 3+0.49[−0.25, +1.23]
Math (40)29 / 9+16.67[+7.50, +26.67]
Physics (20)8 / 3+8.33[−5.00, +23.33]
τ³ Airline (50)39 / 4+23.33[+14.67, +32.67]
τ³ Retail (114)45 / 16+8.48[+3.51, +13.74]

FlowState achieves a higher mean success rate in all six domains. The paired intervals for Shopping, Math, Airline, and Retail lie entirely above zero, providing further evidence of its gains.

AuthorsOpenReview ↗

Response to Qnb9 (2/4): Individual contributions of the design choices

W2: Individual contributions of the design choices

The ablations do not sufficiently connect the overall gains to the distinctive design choices. Removing ISU replaces rich, incrementally maintained states with final-answer nodes created after a request, changing update timing, information granularity, and content simultaneously. Removing PSA exposes all historical states, changing the amount of information in the context. These experiments evaluate useful component-level alternatives, but they do not isolate the benefits of typed states, relations, evidence links, or progressive access.

Thank you for this suggestion. We acknowledge that the original ablations primarily compared components and were insufficient to isolate the contribution of each design choice. To address this concern, we added five controlled variants on Math and Physics. Each starts from the full FlowState framework and changes only one mechanism:

  • No evidence access: The agent can still read historical states but cannot revisit the raw tool outputs supporting them. State content, relations, and current observations are retained.
  • No relation navigation: When a historical state is disclosed, links to its related states are hidden. All historical states remain accessible through the Historical State Index.
  • No node types: We collapse the five state types into a single generic type, keeping all other design choices unchanged.
  • No relation types: We replace the relation labels between states with a single generic type, keeping all other design choices unchanged.
  • Full-state rewriting: We keep the state-update timing unchanged but require the agent to regenerate all modifiable states for the current request, rather than returning only additions, updates, and removals.

SR and PS are reported as percentages, and token consumption is reported in millions. Results are presented as mean ± standard deviation.

VariantMath SRMath PSMath tokens (M)Physics SRPhysics PSPhysics tokens (M)
Full FlowState42.50 ± 5.0055.77 ± 0.2512.95 ± 1.7361.67 ± 2.8974.88 ± 2.101.98 ± 0.18
No evidence access39.17 ± 3.8253.68 ± 1.7014.01 ± 0.4353.33 ± 12.5869.38 ± 3.852.10 ± 0.04
No relation navigation40.83 ± 1.4453.44 ± 1.3813.65 ± 0.1755.00 ± 10.0071.06 ± 5.402.05 ± 0.03
No node types41.67 ± 1.4453.97 ± 2.0013.70 ± 0.0553.33 ± 2.8968.87 ± 3.662.03 ± 0.07
No relation types40.83 ± 1.4455.93 ± 0.6713.87 ± 0.2850.00 ± 5.0069.42 ± 2.172.13 ± 0.06
Full-state rewriting31.67 ± 1.4454.58 ± 1.5413.72 ± 0.1958.33 ± 2.8972.80 ± 2.681.91 ± 0.13

As shown above, all five variants have lower mean SR than full FlowState on both Math and Physics. The first four variants retain access to historical states and test the roles of node types, relation semantics, relation navigation, and access to raw evidence. Full-state rewriting also reduces mean SR despite preserving the update timing and the scope of modifiable states, further supporting the value of incremental maintenance.

Compared with the original component-level ablations, these controlled comparisons connect the overall gains to specific design choices and more directly assess the roles of state representation, historical access, and incremental maintenance. To isolate the benefit of progressive access, we additionally compare PSA with single-pass retrieval using matched historical-state snapshots and exploration input-budget limits (see our response to Q2, “PSA versus single-pass retrieval under matched budgets”). PSA achieves higher mean accuracy in all six settings across two domains and three budget levels, supporting the value of continued, on-demand access to historical states as information needs arise during execution.

AuthorsOpenReview ↗

Response to Qnb9 (3/4): Historical-index growth and context capacity

W3 and Q3: Historical-index growth and context capacity

The implications of historical-index growth remain unclear. The historical index includes summaries from all preceding requests, and disclosed historical states accumulate in the active state. Historical nodes cannot be removed through the same operations available for current-request nodes. Consequently, selective access alone does not establish that the visible context remains compact as the history grows.

How do index size and disclosed historical content grow, and what happens when they exceed the available context capacity?

Thank you for raising this question. In FlowState, the Historical State Index remains in the context, whereas the Active State is specific to the current user request. When a new request arrives, the Active State is reinitialized, and states are disclosed on demand based on the request and the summaries in the Historical State Index. This design keeps the model-visible context compact overall. Our execution logs show that FlowState's context grows substantially more slowly than Full Context as tasks progress.

Context growth during task execution. The figure below compares the context length per LLM call as tasks progress on Math and Physics for FlowState, Full Context, ReasoningBank, BM25, and ZipAct.

https://anonymous.4open.science/api/repo/FlowState-F34D/file/context_growth_300dpi.png

The horizontal axis shows relative task progress, and the vertical axis shows the number of context tokens per LLM call on a logarithmic scale.

In both environments, FlowState's context grows substantially more slowly than Full Context and remains much shorter toward the end of execution. Its context length is also comparable in magnitude to those of ReasoningBank, BM25, and ZipAct. These results show that FlowState can maintain a compact model-visible context while preserving access to historical information.

Growth of the historical index and disclosed historical content. We further measured token growth in the historical index and disclosed historical content on Math and Physics.

https://anonymous.4open.science/api/repo/FlowState-F34D/file/historical_index_disclosed_state_tokens.png

Top row: mean historical-index token count as a function of relative progress across subtasks. Bottom row: mean token count of disclosed historical content as a function of agent-call progress within an individual subtask.

The historical index grows gradually as subtasks progress, reaching mean endpoint sizes of 3.19K tokens on Math and 1.27K tokens on Physics. Disclosed historical content grows slowly within each subtask, reaching mean endpoint sizes of 3.25K and 0.98K tokens, respectively. The Active State is reinitialized at subtask boundaries, so disclosed content does not automatically accumulate across subtasks. These results illustrate how the summary index and on-demand disclosure support both historical reuse and context efficiency.

Context-capacity limits. We acknowledge that the current version does not guarantee that the context always remains within capacity. In extreme cases, FlowState may exceed the available context capacity and cause LLM inference to fail. However, we did not observe such overflows in the experiments reported in this paper. FlowState is compatible with existing memory retrieval and state compression methods, which could be combined with it to address context-capacity overflows. These mechanisms have not yet been implemented in the current version and remain directions for future work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.