Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

SRMT: Shared Memory for Multi-agent Lifelong Pathfinding

T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read A shared memory broadcast lets agents coordinate without explicit messages.

desk verdict A neat shared-memory mechanism for decentralized MARL, but the paper's own RATE-gen control shows initialization is a confound and the abstract overstates the sparse-reward results. read the letter →

arxiv 2501.13200 v1 pith:TSVU3KIF submitted 2025-01-22 cs.LG cs.AIcs.MA

classification cs.LGcs.AIcs.MA
keywords sharedmemorymulti-agentreinforcementlearningpathfindingtransformerrecurrentdecentralizedcoordinationsparserewardPOGEMA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that agents in a decentralized multi-agent system can coordinate simply by pooling their individual recurrent memory states into a shared, globally readable buffer. The proposed Shared Recurrent Memory Transformer (SRMT) gives each agent a personal memory vector, broadcasts all agents' memory vectors as a common sequence each step, and lets every agent cross-attend to that sequence before acting. In a two-agent bottleneck navigation task, SRMT outperforms reinforcement learning baselines, especially when rewards are sparse, and its policies generalize to corridors of length 1000 after training on corridors of length 3 to 30. On the POGEMA benchmark, SRMT is competitive with recent MARL, hybrid, and planning-based methods. A sympathetic reader would care because this suggests that implicit information exchange through a global workspace can replace explicit communication protocols in cooperative multi-agent systems.

What carries the argument

The load-bearing mechanism is the shared memory sequence: at every time step, the memories $\text{mem}_{i,t}$ of all agents $i$ are concatenated into one ordered list, and each agent's transformer performs cross-attention between its own hidden state and that global list. This gives every agent a direct read on the hidden state of every other agent without any explicit message passing, while a memory head rewrites each agent's private memory for the next step. The design extends the Recurrent Memory Transformer to the multi-agent setting, where the shared memory acts as a global workspace consistent with the paper's stated inspiration from global workspace theory.

What would settle it

Train the RMT baseline (recurrent memory without sharing) using SRMT's memory-initialization scheme and compare it with SRMT under Sparse and Moving Negative rewards; if RMT-gen matches SRMT's success rates and corridor-length scaling, the shared-memory cross-attention is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The central discovery claimed is that a shared recurrent memory, implemented as a globally broadcast sequence of per-agent memory vectors, enables implicit coordination in decentralized multi-agent pathfinding. Each agent's policy is a memory transformer: it self-attends over its own recent observations and personal memory, then cross-attends to the shared memory containing every agent's current memory, and updates its personal memory via a memory head. This mechanism is fully decentralized at execution time, requires no hand-designed communication protocol, and maintains a single shared policy across homogeneous agents. The paper reports that SRMT solves the Bottleneck task where other memory and communication baselines fail under sparse reward, and that it remains effective on POGEMA maps including Mazes, Random, MovingAI, and Warehouse when coupled with the Follower heuristic planner.

Load-bearing premise

The paper's headline advantage is attributed to the shared-memory broadcast, but its own ablation shows that simply giving baselines the same memory-initialization scheme as SRMT (as in RATE-gen) closes much of the gap, so the unique contribution of the shared cross-attention is not cleanly isolated.

Editorial extensions

If this is right

  • If SRMT's claims hold, decentralized MARL can achieve coordination through a global memory broadcast rather than learned communication channels, which may simplify training and execution in settings where message protocols are hard to design.
  • SRMT policies trained on short corridors (3–30 cells) would remain effective on corridors up to 1000 cells, suggesting the shared memory encodes reusable coordination heuristics rather than overfitting to a specific corridor length.
  • On the POGEMA benchmark, SRMT would be a competitive option against planning-based centralized methods on throughput, while retaining full decentralization.
  • Combining SRMT with the Follower heuristic planner would yield the best congestion performance in warehouse-like maps, indicating that learned shared memory and classical planning can complement each other.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors' RATE-gen result implies that memory initialization—not shared broadcasting—may be the dominant factor in the Bottleneck gains; a systematic sweep of initialization schemes across all baselines would settle the mechanism.
  • If the shared memory truly carries coordination information, the cosine-distance analysis could be turned into a causal test: permuting or masking the shared memory during evaluation should degrade coordination performance proportionally to the amount of information removed.
  • The shared-memory design may transfer to other cooperative MARL domains beyond pathfinding, such as formation control or task allocation, wherever agents need a cheap global summary of each other's state.
  • Since memory is pooled into a single ordered sequence, the method's scalability depends on the attention cost of reading $n$ memory vectors; the authors' experiments with 64–128 agents suggest an opportunity for a sparse or hierarchical shared memory in larger populations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes Shared Recurrent Memory Transformer (SRMT), a multi-agent architecture in which each agent's recurrent memory vectors are pooled into a shared memory and globally broadcast through cross-attention. The authors evaluate SRMT on a two-agent Bottleneck navigation task under several reward schemes and on the POGEMA lifelong MAPF benchmark. They report that SRMT outperforms existing MARL baselines and ablation variants in the Bottleneck task, especially under sparse rewards, generalizes to corridors longer than those seen in training, and is competitive with recent methods on POGEMA maps. The paper also includes an appendix with additional reward variants, training details, and an analysis of the learned memory representations.

Significance. If the central claim is established, SRMT would be a useful addition to decentralized MARL for pathfinding: it proposes a simple mechanism for implicit information exchange without an explicit communication protocol, and the paper provides code, confidence intervals, and multi-seed evaluations. The authors also provide a useful negative result in the RATE-gen variant, showing that memory initialization alone changes performance substantially. However, the paper's headline claims currently overreach the evidence: the sparse-reward generalization claim is contradicted by Figure 4, and the attribution of SRMT's gains to shared memory is not cleanly isolated from memory-initialization differences. The lifelong MAPF comparison is also weakened by the use of planning-derived dense rewards for SRMT without a stated equivalent for the baselines. These issues are fixable, but they are load-bearing for the paper's main message.

major comments (3)
  1. [Abstract and Section 4.1 (Figure 4)] The abstract claims that SRMT 'consistently outperforms a variety of reinforcement learning baselines, especially under sparse rewards, and generalizes effectively to longer corridors.' This is internally contradicted by the manuscript's own Figure 4 and the text in Section 4.1, which state that for the Sparse reward, SRMT leads only up to corridor length 400 and then drops below RMT for collective performance. The central claim should be qualified to specify the metric, the corridor-length range, and the distinction between individual and cooperative success rates.
  2. [Section 4.1, Figure 4, and Section 3] The RATE-gen experiment demonstrates that memory initialization is a strong confound: re-initializing RATE's memory from the initial observation substantially improves its Moving Negative performance, which the authors attribute to 'the importance of the proper memory initialization procedure.' The paper treats RMT as the key ablation without shared memory, but it never states whether RMT uses the same first-observation-generated initialization as SRMT or a fixed/random vector. If RMT uses a different initialization, the SRMT-vs-RMT gap in the Sparse setting is confounded in exactly the same way. The authors should report RMT's initialization and ideally add an initialization-matched ablation across all architectures, or provide evidence that initialization is not the main driver of the reported gains.
  3. [Section 4.2 (Lifelong MAPF) and Figure 5] In the lifelong experiments, SRMT is trained with a dense reward for following a planned path computed by A* or the Follower heuristic, as described in the paragraph after Figure 5. The comparison with MAMBA and QPLEX is only meaningful if those baselines receive an equivalent reward signal; otherwise, SRMT benefits from auxiliary planning information that the baselines do not have. The paper should state the reward functions used for each baseline and, ideally, include a controlled variant of SRMT trained without the planning-derived reward, or apply the same reward to the baselines.
minor comments (4)
  1. [Section 4.1, Figure 4 caption] The phrase 'collective performance' in the statement about SRMT dropping below RMT is ambiguous; specify whether this refers to CSR, ISR, or SoC.
  2. [Section 3, Figure 1] The text says the input sequence includes 'the historical sequence of the agent's observations from the past h=8 time steps' but it is not clear whether this history is a fixed window of raw observations or processed hidden states; please clarify in the architecture description.
  3. [Appendix A.2, Figure 10] The scalability figure reports shaded confidence intervals, but the number of evaluation runs and seeds is not stated; please provide this detail for reproducibility.
  4. [Section 4.1, Figure 3] The caption lists 'Empty, RNN, Attention' as ablations, but the main text also mentions a GRU-based RNN; please align the naming so that the reader can map each curve to the architecture described in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SRMT is an empirical architecture comparison with self-contained ablations; confounds noted are attribution issues, not circular derivations.

full rationale

The paper does not contain a parametric derivation that reduces to its own output. SRMT is introduced constructively as a policy-core architecture that pools and broadcasts per-agent memory vectors, and the central claims are supported by experiments against baselines implemented in the same framework (RMT, Attention, Empty, RNN as ablations; ATM, RATE, RRNN, MAMBA, QPLEX as external comparisons). The self-citations (Burtsev et al., 2020; Bulatov et al., 2022) are background references for memory transformers; the RMT ablation used in the bottleneck experiments is an in-paper implementation, not an imported uniqueness theorem or fitted ansatz. The authors' candid RATE-gen experiment (Sec. 4.1, Fig. 4) shows that memory initialization is a relevant confound, but reporting a baseline variant is not circular: no fitted value is renamed as a prediction, and the shared-broadcast contribution is not defined in terms of the outcome. The abstract's claim about sparse-reward generalization is somewhat stronger than the Fig. 4 result for corridor lengths beyond 400, but overstatement is a calibration/correctness concern, not a circularity concern. No load-bearing step reduces by construction to its inputs, so the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on standard RL assumptions, on the paper's design choices for memory initialization and planner-based rewards, and on the feasibility of global memory broadcast. The most consequential free parameter is the memory initialization strategy, which the paper itself shows can shift baseline results.

free parameters (4)
  • Learning rate (SRMT) = 0.00013 (MAPF), 0.00022 (LMAPF)
    Selected via grid search, but the reported search range [0.01, 0.05] excludes these values; inconsistency in Appendix A.1.
  • Entropy coefficient (SRMT) = 0.0156 (MAPF), 0.023 (LMAPF)
    Selected via grid search; reported range [0.00001, 0.0003] excludes these values.
  • Discount factor = 0.9716 (MAPF), 0.9756 (LMAPF)
    Chosen for training; affects value learning and credit assignment.
  • Memory initialization strategy = from first observation (SRMT)
    SRMT memory is initialized from first-step observation while baselines use fixed or random values. The RATE-gen experiment shows this choice alone improves baseline performance.
assumptions (4)
  • standard math Reinforcement learning objective, PPO policy-gradient update, and Bellman equations are assumed as background.
    Training uses Sample Factory with PPO; convergence is not proven.
  • domain assumption Agents have perfect localization, synchronized action execution, and static obstacles.
    Stated in the Limitations section; these assumptions exclude perception and motion noise.
  • ad hoc to paper Following a planner-computed path, via A* or the Follower heuristic, is a suitable reward signal for lifelong SRMT.
    Used in Section 4.2 to define r=0.01 rewards; this injects planner knowledge into the learned policy.
  • ad hoc to paper Global broadcasting of all agent memory vectors is an efficient and sufficient coordination mechanism.
    Central to SRMT; no theoretical scaling analysis is given, only empirical results.
invented entities (1)
  • SRMT shared memory block (pooled agent memories with cross-attention) independent evidence
    purpose: Global workspace for implicit inter-agent communication and coordination
    The block is fully specified in Section 3 and its effect is tested against private-memory ablations including RMT, ATM, and RATE, so its contribution has an empirical handle rather than being an unmeasured entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SRMT: Shared Memory for Multi-agent Lifelong Pathfinding." pith.science (2026). https://pith.science/paper/TSVU3KIF

@misc{pith2026250113200,
  author       = {Pith},
  title        = {Pith review of: SRMT: Shared Memory for Multi-agent Lifelong Pathfinding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSVU3KIF}},
  note         = {Machine review of arXiv:2501.13200}
}
read the original abstract

Multi-agent reinforcement learning (MARL) demonstrates significant progress in solving cooperative and competitive multi-agent problems in various environments. One of the principal challenges in MARL is the need for explicit prediction of the agents' behavior to achieve cooperation. To resolve this issue, we propose the Shared Recurrent Memory Transformer (SRMT) which extends memory transformers to multi-agent settings by pooling and globally broadcasting individual working memories, enabling agents to exchange information implicitly and coordinate their actions. We evaluate SRMT on the Partially Observable Multi-Agent Pathfinding problem in a toy Bottleneck navigation task that requires agents to pass through a narrow corridor and on a POGEMA benchmark set of tasks. In the Bottleneck task, SRMT consistently outperforms a variety of reinforcement learning baselines, especially under sparse rewards, and generalizes effectively to longer corridors than those seen during training. On POGEMA maps, including Mazes, Random, and MovingAI, SRMT is competitive with recent MARL, hybrid, and planning-based algorithms. These results suggest that incorporating shared recurrent memory into the transformer-based architectures can enhance coordination in decentralized multi-agent systems. The source code for training and evaluation is available on GitHub: https://github.com/Aloriosa/srmt.

Figures

Figures reproduced from arXiv: 2501.13200 by the authors.

Figure 1
Figure 1. Shared Recurrent Memory Transformer architecture. SRMT pools recurrent mem￾ories memi,t of individual agents at a moment t and provides global access to them via cross￾attention. from a squared window of fixed size centered at the agent’s current location. Formally, the partially observable multi-agent Markov decision process M is defined (Bernstein et al., 2002): M = ⟨S, U, A, P, R, O, O, γ⟩, where S is the set of … view at source ↗
Figure 2
Figure 2. Examples of environments. (a) Bottleneck task. This is a toy task on coordination. Two agents start in rooms opposite their goals and should coordinate passing the corridor. Agents are shown as solid-colored circles, their goals are empty circles with the same border color. (b)-(f) Maps from POGEMA benchmark (images for POGEMA maps are from (Skrynnik et al., 2024a)). POGEMA allows testing the planning methods’ gener… view at source ↗
Figure 3
Figure 3. SRMT effectively solves the Bottleneck Task with different reward functions. Trained with Directional (positive when moved towards a goal and achieved it) reward, SRMT clearly out￾performs the communication (MAMBA, QPLEX) and memory (ATM, RATE, RRNN) baselines. The RMT, Attention, and RNN ablations also solve the task. For the case with the negative reward for movement and no directional reward (Moving Negative) SRM… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: SRMT agents generalize on corridor lengths up to 1000. After training on corridor sizes from 3 to 30 cells all methods were evaluated on longer passages up to 1000. All non-zero performing models show good scaling up to the corridor length of 100. For the Sparse reward…
Figure 5
Figure 5. Figure 5: SRMT outperforms other MARL methods in different environments. SRMT trained on Mazes shows robust generalization when evaluated on maps not seen during training. SRMT out￾performs MARL baselines MAMBA and QPLEX on all maps except the Warehouse environment. Mixed traini…
Figure 6
Figure 6. Figure 6: Comparison of SRMT and other methods across key performance metrics in multi￾agent pathfinding. The bar chart compares the performance of SRMT and its variants (SRMT 64- 128, SRMT-FlwrPlan) against other methods – MAMBA, QPLEX, Follower, MATS-LP, and RHCR – across six …
Figure 7
Figure 7. Figure 7: Trained with Dense reward, all models except empty core policy scale with enlarging corridor length. SRMT consistently outperforms baselines both in success rates and in the time needed to solve the task. The shaded area indicates 95% confidence intervals [PITH_FULL_I…
Figure 8
Figure 8. Figure 8: Directional reward training leads to all the methods preserving the scores for all tested corridor lengths. The shaded area indicates 95% confidence intervals [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Results of training with Directional Negative reward. Vanilla attention fails to scale at corridor lengths of more than 400, compared to the SRMT which preserves the highest scores. That proves the sufficiency of the proposed SRMT architecture. The shaded area indicate…
Figure 10
Figure 10. Figure 10: The evaluation of scalability of SRMT on MovingAI maps from POGEMA benchmark. [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: SRMT memory distances are aligned with the distances between agents. The figure shows how cosine distances between agents’ memory vectors are related to the Euclidean distances between agents on the map for SRMT. The triangle marks the step when agents face each other…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Toward Efficient Agents: Memory, Tool learning, and Planning

    cs.AI 2026-01 conditional novelty 3.0 of 10

    A survey that organizes efficiency techniques for LLM agents into memory, tool learning, and planning, and consolidates benchmarks and metrics for measuring cost-performance trade-offs.

  2. Intelligent System of Emergent Knowledge: A Coordination Fabric for Billions of Minds

    cs.MA 2025-06 reject novelty 2.0 of 10

    ISEK is a conceptual blockchain-and-token architecture for coordinating human and AI agents, with no implementation, experiments, or formal results reported.

Reference graph

Works this paper leans on

16 extracted references · 10 canonical work pages · cited by 2 Pith papers

  1. [1]

    Akshat Agarwal, Sumit Kumar, and Katia P. Sycara. Learning transferable cooperative behavior in multi-agent teams. CoRR, abs/1906.01202,

  2. [8]

    Multi-agent path finding with prioritized communication learning

    Wenhao Li, Hongjun Chen, Bo Jin, Wenzhe Tan, Hong Zha, and Xiangfeng Wang. Multi-agent path finding with prioritized communication learning. 2022 International Conference on Robotics and Automation (ICRA), pp. 10695–10701,

  3. [9]

    Distributed heuristic multi-agent path finding with com- munication

    Ziyuan Ma, Yudong Luo, and Hang Ma. Distributed heuristic multi-agent path finding with com- munication. In 2021 IEEE International Conference on Robotics and Automation (ICRA) , pp. 8699–8705. IEEE, 2021a. Ziyuan Ma, Yudong Luo, and Jia Pan. Learning selective communication for multi-agent path finding. IEEE Robotics and Automation Letters, 7(2):1455–146...

  4. [10]

    Scalable Multi-Agent Reinforcement Learning through Intelligent Information Aggregation

    URL https://arxiv.org/abs/2211.02127. Keisuke Okumura. Lacam: Search-based algorithm for quick multi-agent pathfinding. In Proceed- ings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 11655–11662,

  5. [11]

    Sukhatme, and Vladlen Koltun

    Aleksei Petrenko, Zhehui Huang, Tushar Kumar, Gaurav S. Sukhatme, and Vladlen Koltun. Sam- ple factory: Egocentric 3d control from pixels at 100000 FPS with asynchronous reinforcement learning. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event , volume 119 of Proceedings of Machine Learning ...

  6. [13]

    Pogema: A benchmark platform for cooperative multi-agent navigation, 2024a

    Alexey Skrynnik, Anton Andreychuk, Anatolii Borzilov, Alexander Chernyavskiy, Konstantin Yakovlev, and Aleksandr Panov. Pogema: A benchmark platform for cooperative multi-agent navigation, 2024a. URL https://arxiv.org/abs/2407.14931. Alexey Skrynnik, Anton Andreychuk, Maria Nesterova, Konstantin Yakovlev, and Aleksandr Panov. Learn to follow: Decentralize...

  7. [14]

    Value-decomposition networks for cooperative multi-agent learning based on team reward

    Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinicius Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z Leibo, Karl Tuyls, et al. Value-decomposition networks for cooperative multi-agent learning based on team reward. In Proceedings of the 17th International Conference on Autonomous Agents and MultiAgent Systems, pp...

  8. [1993]

    Reciprocal velocity obstacles for real-time multi-agent navigation

    Jur Van den Berg, Ming Lin, and Dinesh Manocha. Reciprocal velocity obstacles for real-time multi-agent navigation. In Proceedings of The 2008 IEEE International Conference on Robotics and Automation (ICRA 2008), pp. 1928–1935. IEEE,

Show all 16 references
  1. [2016]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun

    URL https://proceedings.neurips.cc/paper_ files/paper/2016/file/c7635bfd99248a2cdef8249ef7bfbef4-Paper.pdf. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. In Proceedings of the IEEE conference on computer vision and patte...

  2. [2017]

    Jianhao Wang, Zhizhou Ren, Terry Liu, Yang Yu, and Chongjie Zhang.{QPLEX}: Duplex dueling multi-agent q-learning

    URL https://proceedings.neurips.cc/paper_files/paper/2017/ file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf. Jianhao Wang, Zhizhou Ren, Terry Liu, Yang Yu, and Chongjie Zhang.{QPLEX}: Duplex dueling multi-agent q-learning. In International Conference on Learning Representations ,

  3. [2018]

    Jiaoyang Li, Andrew Tinka, Scott Kiesel, Joseph W Durham, TK Satish Kumar, and Sven Koenig

    URL http://arxiv.org/abs/1810.02912. Jiaoyang Li, Andrew Tinka, Scott Kiesel, Joseph W Durham, TK Satish Kumar, and Sven Koenig. Lifelong multi-agent path finding in large-scale warehouses. In Proceedings of the AAAI Confer- ence on Artificial Intelligence, volume 35, pp. 11272–11281,

  4. [2019]

    URL http://arxiv.org/abs/1906. 01202. Bernard J. Baars. A Cognitive Theory of Consciousness . Cambridge University Press, New York,

  5. [2020]

    Rae, Mike Chrzanowski, Theophane Weber, Daan Wierstra, Oriol Vinyals, Razvan Pascanu, and Timothy P

    Adam Santoro, Ryan Faulkner, David Raposo, Jack W. Rae, Mike Chrzanowski, Theophane Weber, Daan Wierstra, Oriol Vinyals, Razvan Pascanu, and Timothy P. Lillicrap. Relational recurrent neural networks. CoRR, abs/1806.01822,

  6. [2021]

    Egor Cherepanov, Alexey Staroverov, Dmitry Yudin, Alexey K

    URL https://proceedings.neurips.cc/paper_files/ paper/2021/file/7f489f642a0ddb10272b5c31057f0663-Paper.pdf. Egor Cherepanov, Alexey Staroverov, Dmitry Yudin, Alexey K. Kovalev, and Aleksandr I. Panov. Recurrent action transformer with memory,

  7. [2022]

    Memory transformer

    Mikhail S Burtsev, Yuri Kuratov, Anton Peganov, and Grigory V Sapunov. Memory transformer. arXiv preprint arXiv:2006.11527,

  8. [2023]

    Shariq Iqbal and Fei Sha

    doi: 10.1109/TNNLS.2023.3329530. Shariq Iqbal and Fei Sha. Actor-attention-critic for multi-agent reinforcement learning. CoRR, abs/1810.02912,

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.