Governed Shared Memory for Multi-Agent LLM Systems

Erni Avram; Nurit Cohen-Inger; Oded Margalit; Ran Taig; Yanki Margalit

REVIEW 1 major objections 1 minor 7 cited by

Governed Shared Memory for Multi-Agent LLM Systems

T0 review · 1 major / 1 minor · reviewed 2026-06-25 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Multi-agent LLM systems require explicit governed shared memory abstractions to address four key failure modes that long-context retrieval cannot handle.

desk verdict This paper delivers concrete production metrics on governed multi-agent memory and surfaces real pipeline bugs, but its central claim rests on an unvalidated list of four failure modes. read the letter →

arxiv 2606.24535 v1 pith:EC44P2EL submitted 2026-06-23 cs.AI

Yanki Margalit , Nurit Cohen-Inger , Erni Avram , Ran Taig , Oded Margalit This is my paper · ORCID

classification cs.AI

keywords multi-agentLLMsharedmemorygovernanceprimitivesprovenancetrackingfailuremodesfleetproductionevaluation

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

The paper formalizes the fleet-memory problem for multi-agent LLM environments and pinpoints four foundational failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. It proposes four systems-level primitives—scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation—to mitigate these issues. These are realized in the MemClaw service and tested through the ArgusFleet harness, which reveals both successes in provenance and isolation as well as practical problems in enforcement and pipeline ordering. The work concludes that live evaluation is essential to uncover failures invisible in theoretical designs alone.

What carries the argument

The fleet-memory problem formalized through its four failure modes, addressed by the four primitives of scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation, as implemented in MemClaw and evaluated in ArgusFleet.

What would settle it

Demonstration of a multi-agent LLM fleet using only long-context retrieval that maintains isolation, freshness, consistency, and provenance without the proposed primitives would falsify the necessity of explicit systems-level abstractions.

Watch

Extended reading notes

Core claim

Long-context retrieval alone is insufficient for production multi-agent memory. Governed shared memory demands explicit systems-level abstractions, and live evaluation is vital to expose enforcement and pipeline-ordering failures missed by design-only treatments. The primitives enable 100% provenance reconstruction of derivation chains and zero cross-fleet leakage while optimizing latencies.

Load-bearing premise

That the four failure modes represent the primary and sufficient set of issues that must be addressed for robust fleet memory and that the ArgusFleet harness provides representative coverage of production conditions.

Editorial extensions

If this is right

Provenance tracking successfully reconstructs 100% of depth-four derivation chains with correct writer identity at sub-second per-hop latency.
Policy-governed propagation achieves high intra-fleet visibility with zero cross-fleet leakage.
Strong write mode reduces write-to-visible latency to a single search round-trip.
Live testing uncovers asymmetric scope enforcement where sub-tenant scope was bypassed on direct GET-by-id requests.
Pipeline ordering conflicts can cause premature rejection of contradictory writes by synchronous gates before asynchronous detectors evaluate them.

Reading between the lines

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

The identified failure modes and primitives may generalize to other distributed knowledge systems beyond LLM agents.
Addressing pipeline ordering requires careful design of synchronous and asynchronous components in memory services.
Production services should incorporate live evaluation harnesses like ArgusFleet to validate governance in realistic conditions.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The one thing to know is that this paper takes the fleet-memory problem seriously by defining four failure modes and four primitives, then measuring them in a production service called MemClaw using their ArgusFleet harness. They get solid numbers on provenance and leakage while catching two implementation issues.

They do a good job showing why design-only approaches fall short. The pipeline ordering conflict, where a synchronous gate rejects writes before the async detector runs, is a real example that only shows up in live testing. Reporting the remediation of the scope bypass is also straightforward and helpful.

The soft spot is the assumption that those four modes are the main ones that matter. There's no threat model or empirical check against other potential problems like concurrent consistency or semantic drift, so the necessity of their abstractions isn't fully tested. The evaluation is also self-contained without baselines, which makes it harder to judge how much better this is than simpler methods.

This is for teams building multi-agent LLM applications who care about memory safety in shared environments. It gives practical guidance on what can go wrong and how to instrument for it.

I would send it for peer review. The concrete results and the negative findings are worth referee attention, even with the open question on whether the failure modes are exhaustive.

Referee Report

1 major / 1 minor

Summary. The paper formalizes the fleet-memory problem for multi-agent LLM systems and identifies four foundational failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. It defines four systems-level primitives (scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation) to address them, implements the primitives in the MemClaw production multi-tenant memory service, and evaluates governance properties using the ArgusFleet reproducible harness. The evaluation reports 100% reconstruction of depth-four provenance chains with correct writer identity at sub-second latency, zero cross-fleet leakage, high intra-fleet visibility, and two remediated architectural issues (asymmetric scope enforcement on direct GET-by-id and pipeline ordering conflict between synchronous near-duplicate gates and asynchronous contradiction detection). The central conclusion is that long-context retrieval alone is insufficient and that explicit abstractions plus live evaluation are required to expose enforcement and ordering failures.

Significance. If the results hold, the work offers concrete, production-derived insights into multi-agent memory governance by measuring a live service and disclosing negative findings rather than relying solely on design arguments or simulations. The reproducible ArgusFleet harness and emphasis on pipeline-ordering failures constitute a strength that could guide practical system design in the field.

major comments (1)

[Introduction / fleet-memory problem formalization] Introduction / fleet-memory problem formalization: The four failure modes are presented as foundational and primary without a completeness argument, threat model, or empirical survey establishing that they are the main issues or that other potential problems (e.g., consistency under concurrent agents or cross-model semantic drift) are secondary. The ArgusFleet evaluation tests the implemented primitives on governance dimensions but does not validate whether unaddressed modes would still produce production failures; this assumption is load-bearing for the claim that the four primitives are necessary.

minor comments (1)

[Abstract / Evaluation] Abstract and evaluation section: The reported metrics (100% provenance reconstruction, zero leakage) are given without details on test scale, number of agents/queries, or variance, which would strengthen the reproducibility claim even though the harness itself is described as reproducible.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the fleet-memory problem formalization. We address the major comment below.

read point-by-point responses

Referee: The four failure modes are presented as foundational and primary without a completeness argument, threat model, or empirical survey establishing that they are the main issues or that other potential problems (e.g., consistency under concurrent agents or cross-model semantic drift) are secondary. The ArgusFleet evaluation tests the implemented primitives on governance dimensions but does not validate whether unaddressed modes would still produce production failures; this assumption is load-bearing for the claim that the four primitives are necessary.

Authors: The four failure modes were identified from incidents observed during operation of the MemClaw production service rather than from a formal survey or threat model. The manuscript presents them as foundational in the context of the fleet-memory problem we formalize, but does not assert completeness or that other issues (such as concurrent consistency or cross-model semantic drift) are secondary. The evaluation measures the effectiveness of the four primitives against the modes they target in a live multi-tenant setting; it does not claim to have tested or ruled out unaddressed modes. We will revise the introduction to state explicitly that the modes are derived from production observations, are not asserted to be exhaustive, and that the paper's central claim is the insufficiency of long-context retrieval alone plus the value of live evaluation for exposing enforcement failures. This clarification addresses the scope concern while preserving the reported results. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: paper is implementation and measurement driven with no derivations or self-referential reductions.

full rationale

The paper formalizes the fleet-memory problem by naming four failure modes and defining four primitives to address them, then implements the primitives in MemClaw and measures them via ArgusFleet. No equations, fitted parameters, predictions, or uniqueness theorems appear in the provided text. The identification of failure modes is presented as a modeling premise rather than a derived result, and the evaluation consists of direct runtime measurements (e.g., 100% provenance reconstruction, zero leakage) rather than any quantity that reduces to its own inputs by construction. No self-citations are invoked as load-bearing support for the central claims. This is a standard systems paper whose central claims rest on implementation and live testing, not on circular logic.

Assumptions & free parameters 0 free parameters · 1 assumptions · 3 invented entities

The central claim rests on the domain assumption that the listed failure modes dominate production risks and that the described primitives are both necessary and sufficient; no free parameters or invented physical entities appear.

assumptions (1)

domain assumption Multi-agent LLM environments require robust mechanisms for shared knowledge management.
Opening premise of the abstract that frames the entire contribution.

invented entities (3)

fleet-memory problem
purpose: To name and structure the shared-memory challenges specific to LLM agent fleets.
Newly defined construct that organizes the four failure modes.
MemClaw
purpose: Production implementation of the governance primitives.
The concrete service whose behavior is measured.
ArgusFleet
purpose: Testing harness for the four governance dimensions.
The evaluation framework used to generate the reported results.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Governed Shared Memory for Multi-Agent LLM Systems." pith.science (2026). https://pith.science/paper/EC44P2EL

@misc{pith2026260624535,
  author       = {Pith},
  title        = {Pith review of: Governed Shared Memory for Multi-Agent LLM Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EC44P2EL}},
  note         = {Machine review of arXiv:2606.24535}
}

read the original abstract

Multi-agent LLM environments require robust mechanisms for shared knowledge management. This paper formalizes the fleet-memory problem and identifies four foundational failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. To address these, we define explicit systems-level primitives: scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation. These primitives are implemented in MemClaw, a production multi-tenant memory service, and evaluated via ArgusFleet, a reproducible harness testing four governance dimensions. Rather than a baseline comparison, this study measures a live production service, emphasizing real-world architectural insights and negative results. Key Evaluation Results Provenance: Successfully reconstructed 100% of depth-four derivation chains with correct writer identity at sub-second per-hop latency. Propagation: Demonstrated high intra-fleet visibility with zero cross-fleet leakage. Under strong write mode, write-to-visible latency was optimized to a single search round-trip. Production Architectural Issues Discovered Asymmetric Scope Enforcement: Tenant isolation held, but sub-tenant scope was initially bypassed on direct GET-by-id requests for agent-scoped credentials (disclosed and remediated during the study). Pipeline Ordering Conflict: While contradiction supersession works for admitted writes, a synchronous near-duplicate gate can prematurely reject contradictory writes before the asynchronous contradiction detector can evaluate them. Conclusion: Long-context retrieval alone is insufficient for production multi-agent memory. Governed shared memory demands explicit systems-level abstractions, and live evaluation is vital to expose enforcement and pipeline-ordering failures missed by design-only treatments.

Figures

Figures reproduced from arXiv: 2606.24535 by the authors.

**Figure 1.** System overview. A fleet of cooperating agents writes to and reads from governed shared memory through MemClaw’s [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** The headline finding is that scope enforcement was bimodal at measurement time (the GET-path gap was remediated the next day; §9.1): • On the tenant-key access axis (GET-by-id with the probing tenant in the query string), tenant-key GET exposure was 164/164 = 1.000 (95% Wilson CI [0.977, 1.000]) and the corresponding in-scope miss rate was 0/28 = 0.000. These bulk probes use a tenant-scoped key, which is tenant-wi… view at source ↗

**Figure 2.** Leakage envelope by measurement axis, as measured (2026-05-30). The leak_rate (over expected-deny probes) is the security signal; the miss_rate (over expectedallow probes) is the availability signal. Enforcement was bimodal. The GET-by-id tenant-key axis returned every requested row to a tenant-scoped caller (tenant-key exposure = 1.000), which is expected under tenant-wide authority. The sub-tenant enforcement g… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Per-hop fetch-latency distribution across all provenance chain walks. The p50 of 291 ms and p95 of 491 ms support chain reconstruction at interactive latencies even at depth four (p99 = 1.1 s). 8.3 Propagation All 40 planned writes landed and all 200 downstream visibi…

**Figure 5.** Figure 5: Visibility rate by reader relation. Fleet siblings [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Write-to-visible window from the dedicated win [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

**Figure 7.** Figure 7: Write-latency distribution under write_mode = strong for the contradiction experiment, in milliseconds (p50 1,840, p95 4,861; dashed/dotted markers). Contradiction detection runs post-commit and asynchronously, so it contributes nothing to this synchronous write laten…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

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

MemTX: Transactional Belief Commit for Stateful Agent Memory
cs.AI 2026-07 conditional novelty 7.0 of 10

Staging agent-memory writes through a validate-and-commit pipeline with maturity-gated irreversible actions and typed cascading repair yields zero realized downstream harm on five LLM backbones, where eight baselines ...
Beyond Agent Harnesses: Cross-Substrate Authority for Multi-Agent Systems
cs.MA 2026-09 accept novelty 6.0 of 10

A controlled experimental benchmark shows that raw authority receipts resolve multi-agent safety decisions, while execution-time enforcement prevents unsafe intents that planning alone cannot.
Counter-Swarm Doctrine: Containing Coordinated Agent Intrusions
cs.CR 2026-09 conditional novelty 6.0 of 10

Defenders should organize agent-intrusion response around revisable, policy-labelled coordination episodes discovered from shared-state evidence, an approach this paper defines and designs an evaluation for but does n...
SoK: When Safe Agents Fail Together: The Security of Multi Agent LLM Systems
cs.CR 2026-09 accept novelty 6.0 of 10

A systematization of security in multi-agent LLM systems introducing an execution-centered A-I-R framework that traces adversarial influence through interaction interfaces to system-level risks.
CrystalMem: Elastic Memory for Self-Evolving LLM Agents via Knowledge Crystallization
cs.AI 2026-07 conditional novelty 6.0 of 10

CrystalMem restores agent capability after a memory-budget squeeze-and-recovery cycle to within 0.5 pp on average by keeping recoverable residue of demoted entries and verifying regenerated knowledge.
SuperLocalMemory 4.0: The Governed Memory Operating System for AI Agents
cs.AI 2026-08 conditional novelty 5.0 of 10

SuperLocalMemory 4.0 is a local-first, governed memory runtime for AI agents with a verifiable transaction spine, cross-store erasure receipts, and measured control-plane overhead.
Always-OnAgents:A Survey of Persistent Memory, State, and Governance in LLMAgents
cs.MA 2026-06 unverdicted novelty 5.0 of 10

Survey mapping persistent state in LLM agents along six axes and proposing the AOEP-v0 protocol to evaluate governance and recovery obligations.

Reference graph

Works this paper leans on

25 extracted references · 19 canonical work pages · cited by 7 Pith papers

[1]

Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory

Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh,andDeshrajYadav. Mem0: Buildingproduction- readyAIagentswithscalablelong-termmemory,2025. URLhttps://arxiv.org/abs/2504.19413

work page Pith review arXiv 2025
[2]

Corbett, Jeffrey Dean, Michael Epstein, An- drew Fikes, Christopher Frost, J

James C. Corbett, Jeffrey Dean, Michael Epstein, An- drew Fikes, Christopher Frost, J. J. Furman, Sanjay Ghemawat, Andrey Gubarev, Christopher Heiser, Pe- ter Hochschild, Wilson Hsieh, Sebastian Kanthak, Eu- gene Kogan, Hongyi Li, Alexander Lloyd, Sergey Mel- nik, David Mwaura, David Nagle, Sean Quinlan, Ra- jesh Rao, Lindsay Rolig, Yasushi Saito, Michal ...
[3]

Corbett, Jeffrey Dean, Michael Epstein, Andrew Fikes, Christopher Frost, J

ISSN 0734-2071. doi: 10.1145/2491245. URL https://doi.org/10.1145/2491245

work page doi:10.1145/2491245
[4]

Securing AI Agents with Information-Flow Control

Manuel Costa, Boris Köpf, Aashish Kolluri, Andrew Paverd,MarkRussinovich,AhmedSalem,ShrutiTople, Lukas Wutschitz, and Santiago Zanella-Béguelin. Se- curing AI agents with information-flow control, 2025. URLhttps://arxiv.org/abs/2505.23643

work page Pith review arXiv 2025
[5]

Defeating Prompt Injections by Design

EdoardoDebenedetti,IliaShumailov,TianqiFan,Jamie Hayes, Nicholas Carlini, Daniel Fabian, Christoph Kern, Chongyang Shi, Andreas Terzis, and Flo- rian Tramèr. Defeating prompt injections by design. arXiv:2503.18813, 2025

work page Pith review arXiv 2025
[6]

Dynamo: Amazon’s highly avail- able key-value store.ACM SIGOPS Operating Systems 14 Review, 41(6):205–220, 2007

Giuseppe DeCandia, Deniz Hastorun, Madan Jampani, Gunavardhan Kakulapati, Avinash Lakshman, Alex Pilchin,SwaminathanSivasubramanian,PeterVosshall, and Werner Vogels. Dynamo: Amazon’s highly avail- able key-value store.ACM SIGOPS Operating Systems 14 Review, 41(6):205–220, 2007. doi: 10.1145/1294261. 1294281

work page doi:10.1145/1294261 2007
[7]

Hu, David Ferraiolo, Rick Kuhn, Adam Schnitzer, Kenneth Sandlin, Robert Miller, and Karen Scarfone

Vincent C. Hu, David Ferraiolo, Rick Kuhn, Adam Schnitzer, Kenneth Sandlin, Robert Miller, and Karen Scarfone. Guide to attribute based access control (ABAC) definition and considerations. NIST Special Publication800-162,NationalInstituteofStandardsand Technology, 2014

2014
[8]

Time, clocks, and the ordering of events in a distributed system

LeslieLamport.Time,clocks,andtheorderingofevents in a distributed system.Communications of the ACM, 21(7):558–565, 1978. doi: 10.1145/359545.359563

work page doi:10.1145/359545.359563 1978

Show all 25 references

[9]

LangMem: Long-term memory for LLM agents.https://langchain-ai.github.io/ langmem/, 2024

LangChain. LangMem: Long-term memory for LLM agents.https://langchain-ai.github.io/ langmem/, 2024

2024
[10]

Retrieval-augmented generation for knowledge-intensive NLP tasks

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler,MikeLewis,WentauYih,TimRocktäschel,Se- bastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Infor...

2020
[11]

A comprehensive sur- vey on long context language modeling, 2025

Jiaheng Liu, Dawei Zhu, Zhiqi Bai, Yancheng He, Huanxuan Liao, Haoran Que, Zekun Wang, Chenchen Zhang,GeZhang,JiebinZhang,YuanxingZhang,Zhuo Chen, Hangyu Guo, Shilong Li, Ziqiang Liu, Yong Shan, Yifan Song, Jiayi Tian, Wenhao Wu, Zhejian Zhou, Ruijie Zhu, Junlan Feng, Yang Gao...

2025
[12]

PeerRank: Autonomous LLM evaluation through web-grounded, bias-controlled peer review, 2026

Yanki Margalit, Erni Avram, Ran Taig, Oded Margalit, and Nurit Cohen-Inger. PeerRank: Autonomous LLM evaluation through web-grounded, bias-controlled peer review, 2026. URLhttps://arxiv.org/abs/2602. 02589

2026
[13]

CanLLMskeepasecret? testingprivacyimplica- tionsoflanguagemodelsviacontextualintegritytheory,

Niloofar Mireshghallah, Hyunwoo Kim, Xuhui Zhou, Yulia Tsvetkov, Maarten Sap, Reza Shokri, and Yejin Choi. CanLLMskeepasecret? testingprivacyimplica- tionsoflanguagemodelsviacontextualintegritytheory,
[14]

URLhttps://arxiv.org/abs/2310.17884
[15]

Patil, Ion Stoica, and Joseph E

CharlesPacker,SarahWooders,KevinLin,VivianFang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonza- lez. MemGPT: Towards LLMs as operating systems. arXiv:2310.08560, 2023

2023 arXiv
[16]

Zep: A tem- poral knowledge graph architecture for agent memory

Preston Rasmussen, Pavlo Paliychuk, Travis Beau- vais, Jack Ryan, and Daniel Chalef. Zep: A tem- poral knowledge graph architecture for agent memory. arXiv:2501.13956, 2025

2025 arXiv
[17]

Collaborativememory: Multi- user memory sharing in LLM agents with dynamic ac- cess control

AlirezaRezazadeh,ZichaoLi,AngeLou,YuyingZhao, WeiWei,andYujiaBao. Collaborativememory: Multi- user memory sharing in LLM agents with dynamic ac- cess control. arXiv:2505.18279, 2025

2025
[18]

RaviS.Sandhu,EdwardJ.Coyne,HalL.Feinstein,and Charles E. Youman. Role-based access control models. IEEE Computer, 29(2):38–47, 1996. doi: 10.1109/2. 485845

1996 doi
[19]

Conflict-free replicated data types

Marc Shapiro, Nuno Preguiça, Carlos Baquero, and Marek Zawirski. Conflict-free replicated data types. In Symposium on Self-Stabilizing Systems (SSS), volume 6976ofLectureNotesinComputerScience,pages386– 400.Springer,2011.doi: 10.1007/978-3-642-24550-3_ 29

2011 doi
[20]

Terry, Marvin M

Douglas B. Terry, Marvin M. Theimer, Karin Petersen, Alan J. Demers, Mike J. Spreitzer, and Carl H. Hauser. ManagingupdateconflictsinBayou,aweaklyconnected replicatedstoragesystem. InACMSymposiumonOper- atingSystemsPrinciples(SOSP),pages172–182,1995. doi: 10.1145/224056.224070

1995 doi
[21]

Unveiling privacy risks in LLM agent memory, 2025

Bo Wang, Weiyi He, Shenglai Zeng, Zhen Xiang, Yue Xing, Jiliang Tang, and Pengfei He. Unveiling privacy risks in LLM agent memory, 2025. URLhttps:// arxiv.org/abs/2502.13172

2025
[22]

MIRIX: Multi-agent memory systemforLLM-basedagents

Yu Wang and Xi Chen. MIRIX: Multi-agent memory systemforLLM-basedagents. arXiv:2507.07957,2025

2025 arXiv
[23]

Auto- Gen: Enabling next-gen LLM applications via multi- agentconversation,2023

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. Auto- Gen: Enabling next-gen LLM applications via multi- agentconversation,2023. URLhttps...

2023 arXiv
[24]

A-MEM: Agentic memory for LLM agents

Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-MEM: Agentic memory for LLM agents. arXiv:2502.12110, 2025

2025 arXiv
[25]

URL https://arxiv.org/abs/2506.07398

Guibin Zhang, Muxin Fu, Guancheng Wan, Miao Yu, KunWang,andShuichengYan.G-Memory: Tracinghi- erarchicalmemoryformulti-agentsystems,2025. URL https://arxiv.org/abs/2506.07398. 15

2025

Pith tools

Reviewed June 25, 2026 · model on record in the stance chip above.

[1] [1]

Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory

Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh,andDeshrajYadav. Mem0: Buildingproduction- readyAIagentswithscalablelong-termmemory,2025. URLhttps://arxiv.org/abs/2504.19413

work page Pith review arXiv 2025

[2] [2]

Corbett, Jeffrey Dean, Michael Epstein, An- drew Fikes, Christopher Frost, J

James C. Corbett, Jeffrey Dean, Michael Epstein, An- drew Fikes, Christopher Frost, J. J. Furman, Sanjay Ghemawat, Andrey Gubarev, Christopher Heiser, Pe- ter Hochschild, Wilson Hsieh, Sebastian Kanthak, Eu- gene Kogan, Hongyi Li, Alexander Lloyd, Sergey Mel- nik, David Mwaura, David Nagle, Sean Quinlan, Ra- jesh Rao, Lindsay Rolig, Yasushi Saito, Michal ...

[3] [3]

Corbett, Jeffrey Dean, Michael Epstein, Andrew Fikes, Christopher Frost, J

ISSN 0734-2071. doi: 10.1145/2491245. URL https://doi.org/10.1145/2491245

work page doi:10.1145/2491245

[4] [4]

Securing AI Agents with Information-Flow Control

Manuel Costa, Boris Köpf, Aashish Kolluri, Andrew Paverd,MarkRussinovich,AhmedSalem,ShrutiTople, Lukas Wutschitz, and Santiago Zanella-Béguelin. Se- curing AI agents with information-flow control, 2025. URLhttps://arxiv.org/abs/2505.23643

work page Pith review arXiv 2025

[5] [5]

Defeating Prompt Injections by Design

EdoardoDebenedetti,IliaShumailov,TianqiFan,Jamie Hayes, Nicholas Carlini, Daniel Fabian, Christoph Kern, Chongyang Shi, Andreas Terzis, and Flo- rian Tramèr. Defeating prompt injections by design. arXiv:2503.18813, 2025

work page Pith review arXiv 2025

[6] [6]

Dynamo: Amazon’s highly avail- able key-value store.ACM SIGOPS Operating Systems 14 Review, 41(6):205–220, 2007

Giuseppe DeCandia, Deniz Hastorun, Madan Jampani, Gunavardhan Kakulapati, Avinash Lakshman, Alex Pilchin,SwaminathanSivasubramanian,PeterVosshall, and Werner Vogels. Dynamo: Amazon’s highly avail- able key-value store.ACM SIGOPS Operating Systems 14 Review, 41(6):205–220, 2007. doi: 10.1145/1294261. 1294281

work page doi:10.1145/1294261 2007

[7] [7]

Hu, David Ferraiolo, Rick Kuhn, Adam Schnitzer, Kenneth Sandlin, Robert Miller, and Karen Scarfone

Vincent C. Hu, David Ferraiolo, Rick Kuhn, Adam Schnitzer, Kenneth Sandlin, Robert Miller, and Karen Scarfone. Guide to attribute based access control (ABAC) definition and considerations. NIST Special Publication800-162,NationalInstituteofStandardsand Technology, 2014

2014

[8] [8]

Time, clocks, and the ordering of events in a distributed system

LeslieLamport.Time,clocks,andtheorderingofevents in a distributed system.Communications of the ACM, 21(7):558–565, 1978. doi: 10.1145/359545.359563

work page doi:10.1145/359545.359563 1978