Memory as a Service (MaaS): Purpose-Bound Memory Mediation for Cooperative Agents

Haichang Li

REVIEW 4 major objections 4 minor 27 references

Memory as a Service (MaaS): Purpose-Bound Memory Mediation for Cooperative Agents

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Agent memory should be a callable service whose every use is mediated by owner, requester, recipient, task, and declared purpose.

desk verdict A coherent position paper that organizes the agent-memory design space well, but the metadata abstract claims experiments the body doesn't contain and the routing layer's core guarantee is unvalidated. read the letter →

arxiv 2506.22815 v2 pith:4QIJY77O submitted 2025-06-28 cs.HC

Haichang Li This is my paper

classification cs.HC

keywords MemoryasaServiceLLMagentscontextualpurpose-boundmediationgovernancecross-entitycollaborationprivacyagentsharing

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 position paper argues that memory in LLM-based agent systems should be redesigned as a service rather than local state. It proposes Memory as a Service (MaaS), where each memory module is independently callable, dynamically composable, and governable, and where every invocation is mediated by owner, requester, recipient, task, and declared purpose. The mediator chooses to withhold, abstract, or reveal each candidate memory item, so private memory can serve public, cross-entity needs without sacrificing ownership and control. To ground the position, the paper reports diagnostic stress-test numbers: relevance-based retrieval is near chance at separating private from shareable items (AUROC 0.570) and leaks 53.0% of private items, while contextual-integrity prompting cuts leakage by 21.8 percentage points yet still leaves 32.6% residual leakage, motivating memory governance as a separate design problem.

What carries the argument

The argument is carried by a three-element mediation decision—withhold, abstract, reveal—issued by a Memory Routing Layer for each candidate memory item, based on owner, requester, recipient, task, and declared purpose. The paper formalizes the mediation space by separating cooperative utility (how much the item helps the requester's task), disclosure leakage (how much private content escapes), and purpose-bound authorization (whether the declared purpose justifies access). Memory Containers package the data together with access-policy metadata, so governance is embedded in the asset itself; the routing layer acts as the trusted environment that interprets requests and enforces policy.

What would settle it

A decisive test would be to build a purpose-bound mediator and run it on a corpus where owners mark each item as shareable or private for each requester-purpose pair; if any reveal exposes an owner-marked-private item, or any withhold blocks an owner-approved one, the claim that intent can be inferred and enforced fails. The paper's own residual-leakage number of 32.6% under contextual-integrity prompting already provides a partial check along these lines.

Watch

Extended reading notes

Core claim

Memory as a Service is the claim that contextual memory—the dynamic external information tied to an entity's experience—should be decoupled from the entity, session, or task that produced it and exposed as a modular service with an addressable endpoint. The paper's central device is purpose-bound mediation: for each invocation, the system evaluates who owns the memory, who is requesting it, who will receive it, the task at hand, and the declared purpose, then selects one of three actions—withhold, abstract, or reveal—for each candidate item. This, the paper argues, resolves the tension between memory's inherently private nature and its potential public value: owners keep sovereign control encoded in memory containers and permission policies, while a memory routing layer provides a standardized public channel for authorized circulation. The position is supported by diagnostic stress tests showing that ordinary relevance retrieval does not respect disclosure boundaries, that purpose-aware prompting reduces but does not eliminate leakage, and that some private items already carry explicit safe-hint abstractions, indicating that memory governance deserves to be studied as its own design problem.

Load-bearing premise

The whole framework depends on an intelligent routing layer that can read the true intent of every natural-language request and enforce the memory owner's declared purpose without leaking private content; no implementation or benchmark shows this is possible.

Editorial extensions

If this is right

Within a single user's ecosystem, memory modules can be shared across persona-based agents while each persona's module stays sealed from the others.
Across organizations, a group can publish an authoritative memory module that members subscribe to, while each member's interactions with it accumulate in a private container.
Legacy memory becomes feasible: a person can package structured memories with revocation rules so a digital persona can interact with loved ones under constrained, revocable access.
Memory markets would be a natural consequence, with high-quality modules treated as priced assets usable through subscriptions and contribution-based dividends.
Governance moves out of application code and into the memory asset itself, making the mediation layer the natural place for auditing and provenance.

Reading between the lines

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

A testable extension the paper leaves open: build a benchmark where owners pre-label ground-truth withhold/abstract/reveal decisions and evaluate a routing layer on both privacy and task utility, not just leakage.
The reported leakage numbers imply a utility-privacy trade-off, but the paper does not measure it; an experiment could quantify how much task performance falls as leakage is driven down.
The purpose-bound framing should generalize to human-facing personal data sharing, where appropriateness depends on relationship and purpose, not just on the requester's identity.

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.

Desk Editor's Note

read the letter

This is a useful position paper, not a research paper, and the metadata doesn't match the manuscript. The arXiv abstract claims diagnostic stress tests on MAGPIE (AUROC 0.570, 53% leakage, 4.5% safe hints) and titles the paper 'Purpose-Bound Memory Mediation for Cooperative Agents,' but the full text contains no MAGPIE, no methodology, no numbers, and a different title. That mismatch is not cosmetic; it reports empirical claims a reviewer cannot check, and it has to be resolved before anything else.

What the paper actually does well: it is honest about its scope. The conclusion says MaaS is 'not a novel invention but rather a systematization and theoretical elevation of existing explorations,' and that is exactly what it is. The two-dimensional design space (entity structure by service type) is a genuinely useful organizing lens, and treating intra-, inter-, and group-level memory as one governed service continuum is a clean way to unify MemGPT, Mem0, MCP, A2A, Collaborative Memory, Factored Agents, and Second Me. The citation pattern is fine; the MemOS reference is a different group's work, so there is no self-citation problem. The writing is clear.

The soft spots, in proportion. First, the metadata mismatch means the only empirical support for the framework's central promise is either missing or mislabeled. Second, the load-bearing component is the Memory Routing Layer's intent-aware enforcement of withhold/abstract/reveal, and it is both unvalidated and underspecified. Section 3.1 says authorization relies on in-context learning during routing, but there is no account of how purpose is represented or validated, or how a legitimate request is distinguished from a prompt-injection or purpose-spoofing attempt. Section 4.2 designates the routing layer as the trusted execution environment, while Section 5.2 concedes that memory pollution and spurious injection are open problems. So the framework's core security promise rests on the least specified part of the architecture, and the paper itself says the framework is at a 'preliminary conceptual stage.' None of this is fatal for a position paper, but it should be unambiguous that this is a framework proposal with no empirical validation.

This is for researchers designing memory architectures for LLM agents, and for survey-style work on multi-agent memory. It deserves a serious referee as a position paper. The conditions I'd attach: align the abstract with the body, and either include the MAGPIE experiments with methodology or drop the numbers. A small prototype or benchmark would materially raise its value, but it is worth engaging with even without one.

Referee Report

4 major / 4 minor

Summary. This paper proposes "Memory as a Service" (MaaS), a conceptual framework in which contextual memory is decoupled from individual agents and exposed as independently addressable, composable, and governable service modules. The framework introduces Memory Containers for packaging private memory with policy metadata and a Memory Routing Layer that adjudicates requests by inferring the owner, requester, recipient, task, and declared purpose, choosing among withhold, abstract, or reveal for each candidate item. The authors position MaaS as a generalization of existing intra-entity memory systems and illustrate its application through a two-dimensional design space (entity structure and service type), followed by an open research agenda on governance, security, and ethics. The body of the paper is conceptual and contains no implementation, formal analysis, or empirical evaluation; the metadata abstract, however, reports numerical diagnostic results on the MAGPIE benchmark that do not appear anywhere in the full text.

Significance. The paper addresses a timely and real problem: enabling cross-entity memory sharing without sacrificing privacy. Its conceptual contribution is a useful systematization of existing memory practices, and the proposed design space (intra-, inter-, and group-level × injective/exchange-based) provides a clear vocabulary for discussing memory governance. The paper also responsibly cites prior work and identifies several important open problems. However, the central privacy guarantee rests on an unvalidated assumption that an LLM-based Memory Routing Layer can reliably infer and enforce purpose and intent. No evidence is provided for this assumption, and the empirical results claimed in the abstract are absent from the manuscript. As a position paper, the framework has merit, but as submitted the abstract overstates the evidential basis and the core mechanism remains unsubstantiated.

major comments (4)

[Abstract (metadata) vs. Full Text] The submitted abstract reports specific empirical results on MAGPIE: relevance-based retrieval reaches AUROC 0.570 and leaks 53.0% of private items, contextual-integrity prompting reduces leakage by 21.8 percentage points to 32.6% residual, and 4.5% of private items contain safe-hint abstractions. None of these numbers, the MAGPIE dataset, the experimental protocol, or any analysis appear in the full text. The full-text abstract is a different, purely conceptual abstract. This is a load-bearing mismatch: either the experiments must be included with sufficient detail to be evaluated, or all empirical claims must be removed from the metadata abstract. As written, the paper contains unsupported empirical assertions.
[Section 3.2] The entire MaaS value proposition depends on the Memory Routing Layer correctly inferring the purpose and intent of each request and reliably choosing withhold/abstract/reveal. The paper describes this layer as performing 'semantic or goal-oriented analysis' but provides no specification of how purpose is represented, validated, or bounded, and no demonstration that an LLM can perform this adjudication without leaking private content. Given the abstract's own admission that even contextual-integrity prompting leaves 32.6% residual leakage, the core inferential step is unproven. If the routing layer cannot reliably distinguish a legitimate request from a purpose-spoofing or prompt-injection attack, the central claim of safe, purpose-bound mediation collapses.
[Sections 4.2 and 5.2] Section 4.2 designates the Memory Routing Layer as the trusted execution environment for exchange-based services, making the system's privacy guarantees depend on the unvalidated routing mechanism. Section 5.2, however, lists 'spurious memory injection', provenance verification, and bias inheritance as open problems, which concedes that the security property is not yet solved. The paper should acknowledge this tension explicitly: the architecture currently places trust in a component whose core capability is an open research question. This should be framed as a hypothesis requiring validation, not as a working design.
[Abstract (metadata) and Sections 1, 6] The metadata abstract states that the paper 'formalizes' the separation of cooperative utility, disclosure leakage, and purpose-bound authorization, but the full text contains no equations, definitions, or formal statements of these quantities. The body is a qualitative position statement; the Conclusion even calls the framework 'preliminary conceptual stage.' The claim of formalization is therefore not supported by the manuscript and should either be substantiated with a formal model or removed from the abstract.

minor comments (4)

[Section 5.2 heading] The heading 'Security and T rust' contains a line-break typo: 'T rust' should be 'Trust'.
[Abstract vs. Full Text] The metadata abstract and the full-text abstract describe different papers; the authors should reconcile them so that the abstract accurately reflects the content and scope of the manuscript.
[Section 2.1] The sentence 'MemOS Li et al. (2025)' is missing proper punctuation ('MemOS (Li et al., 2025)') and should be corrected for clarity.
[Section 4.2] The phrase 'temporary, neutral trusted execution environment' is used informally; if a formal notion of trusted execution is intended, it should be defined or the informal usage should be flagged as an analogy.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MaaS is an unvalidated position/taxonomy paper; missing MAGPIE support and routing-layer assumptions are evidence gaps, not circular reductions.

full rationale

This paper contains no empirical prediction or first-principles derivation whose output is equivalent to its input. It has no equations, no fitted parameters, and no benchmark results in the body to reduce. The abstract claims a formalization and MAGPIE diagnostic stress tests, but those are absent from the full text; that is an omitted-support/verifiability problem, not circularity. The conclusion explicitly concedes the framework is "still in its preliminary conceptual stage" (Section 6), and Section 5.2 lists provenance and spurious-memory injection as open problems, acknowledging that the Memory Routing Layer's core security property is unresolved. The load-bearing assumption that an LLM routing layer can infer purpose and enforce withhold/abstract/reveal (Sections 3.1, 3.2, 4.2) is asserted rather than derived, and an unvalidated assumption is not the same as a circular derivation. The paper also does not rely on load-bearing self-citations: the only "Li et al. 2025" reference is MemOS by different authors, and no uniqueness theorem is imported from prior work. The design-space classification of existing memory systems is explicitly framed as a "systematization and theoretical elevation" rather than a prediction from first principles, so no circular step can be exhibited. The main risks are correctness and completeness risks, not circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 2 invented entities

The paper introduces no fitted parameters. Its central claim rests on conceptual premises about the modularizability and governability of memory, none of which are demonstrated. The two invented architectural components, Memory Container and Memory Routing Layer, have no external evidence and no falsifiable implications in the text.

assumptions (3)

domain assumption Contextual memory can be cleanly separated from parametric memory and treated as an independently addressable module.
Section 3 states that contextual memory is external information tied to an entity's experience and can be modularized, but this modularity is assumed without evidence.
domain assumption Memory is a data asset, so ownership and control can be attached to it.
Section 3.2 cites the NIST definition of data asset to ground memory ownership, an interpretive premise that drives the whole governance design.
ad hoc to paper An LLM-based routing layer can infer and honor declared purpose and intent to decide withhold/abstract/reveal.
Section 3.2 proposes the Memory Routing Layer without implementation or evaluation in the body; the metadata abstract's MAGPIE results are not present in the text.

invented entities (2)

Memory Container
purpose: Encapsulates memory content plus access policy metadata to embed governance directly into the asset.
Proposed in Section 3.2; no implementation, API specification, or falsifiable handle is provided.
Memory Routing Layer
purpose: Adjudicates and routes service requests, deciding withhold/abstract/reveal based on owner, requester, recipient, task, and purpose.
Proposed in Section 3.2; no prototype or benchmark exists in the body, and the abstract's empirical results are absent from the text.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Memory as a Service (MaaS): Purpose-Bound Memory Mediation for Cooperative Agents." pith.science (2026). https://pith.science/paper/4QIJY77O

@misc{pith2026250622815,
  author       = {Pith},
  title        = {Pith review of: Memory as a Service (MaaS): Purpose-Bound Memory Mediation for Cooperative Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4QIJY77O}},
  note         = {Machine review of arXiv:2506.22815}
}

abstract

Agentic programming is code-centered, while its useful memory context extends beyond code. A programming agent may draw on memory from test, review, build, and release agents; design, product, security, operations, and compliance agents; meeting, finance, calendar, and workflow agents; personal agents; and agents acting for other people. These memories can help agents optimize, debug, test, and evaluate software, while carrying different owners, purposes, recipients, and disclosure boundaries. We propose \emph{Memory as a Service} (MaaS) as \emph{purpose-bound memory mediation}: each invocation is evaluated by owner, requester, recipient, task, and declared purpose, and the mediator chooses whether to \emph{withhold}, \emph{abstract}, or \emph{reveal} each candidate item. We formalize this by separating cooperative utility, disclosure leakage, and purpose-bound authorization, then ground the position with diagnostic stress tests on MAGPIE. Relevance-based retrieval reaches AUROC $0.570$ and leaks $53.0\%$ of private items; contextual-integrity prompting reduces leakage by $21.8$ percentage points while leaving $32.6\%$ residual leakage; and $4.5\%$ of private items contain explicit safe-hint abstractions. These probes motivate memory governance as a separate design problem for cooperative programming agents.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 10 canonical work pages

[1]

Model context protocol

Anthropic . Model context protocol. GitHub repository, 2024. URL https://github.com/modelcontextprotocol. A protocol for connecting AI assistants to data sources and tools

work page 2024
[2]

Anthropic . Claude. https://claude.ai, 2025. Accessed: June 27, 2025

work page 2025
[3]

Mem0: Building production-ready ai agents with scalable long-term memory

Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:2504.19413, 2025

arXiv 2025
[4]

Yiming Du, Wenyu Huang, Danna Zheng, Zhaowei Wang, Sebastien Montella, Mirella Lapata, Kam-Fai Wong, and Jeff Z. Pan. Rethinking memory in ai: Taxonomy, operations, topics, and future directions, 2025. URL https://arxiv.org/abs/2505.00675

arXiv 2025
[5]

Memory sharing for large language model based agents, 2024

Hang Gao and Yongfeng Zhang. Memory sharing for large language model based agents, 2024. URL https://arxiv.org/abs/2404.09982

arXiv 2024
[6]

Agent2agent (a2a) protocol

Google. Agent2agent (a2a) protocol. GitHub repository, 2024. URL https://github.com/a2aproject/A2A. Open-source framework for multi-agent systems

work page 2024
[7]

Google . Gemini. https://gemini.google.com, 2025. Accessed: June 27, 2025

work page 2025
[8]

Letters from future self: Augmenting the letter-exchange exercise with llm-based agents to enhance young adults' career exploration

Hayeon Jeon, Suhwoo Yoon, Keyeun Lee, Seo Hyeong Kim, Esther Hehsun Kim, Seonghye Cho, Yena Ko, Soeun Yang, Laura Dabbish, John Zimmerman, Eun-mee Kim, and Hajin Lim. Letters from future self: Augmenting the letter-exchange exercise with llm-based agents to enhance young adults' career exploration. In Proceedings of the 2025 CHI Conference on Human Factor...

arXiv 2025

Show all 27 references

[9]

Personalized federated learning techniques: Empirical analysis, 2024

Azal Ahmad Khan, Ahmad Faraz Khan, Haider Ali, and Ali Anwar. Personalized federated learning techniques: Empirical analysis, 2024. URL https://arxiv.org/abs/2409.06805

2024 arXiv
[10]

Memos: An operating system for memory-augmented generation (mag) in large language models, 2025

Zhiyu Li, Shichao Song, Hanyu Wang, Simin Niu, Ding Chen, Jiawei Yang, Chenyang Xi, Huayi Lai, Jihao Zhao, Yezhaohui Wang, Junpeng Ren, Zehao Lin, Jiahao Huo, Tianyi Chen, Kai Chen, Kehang Li, Zhiqiang Yin, Qingchen Yu, Bo Tang, Hongkang Yang, Zhi-Qin John Xu, and Feiyu Xiong....

2025 arXiv
[11]

The New Information Goldmine

Dyan Machan. The New Information Goldmine . The Wall Street Journal, aug 2009. URL https://www.wsj.com/articles/SB125071202052143965. Retrieved June 9, 2010

2009
[12]

Openmemory

Mem0 . Openmemory. https://docs.mem0.ai/openmemory/overview, 2025. Accessed: June 27, 2025

2025
[13]

Memgraph: Real-time graph streaming platform

Memgraph Ltd. Memgraph: Real-time graph streaming platform. https://memgraph.com, 2025. Accessed: June 27, 2025

2025
[14]

Brubaker

Meredith Ringel Morris and Jed R. Brubaker. Generative ghosts: Anticipating benefits and risks of ai afterlives. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI '25, New York, NY, USA, 2025. Association for Computing Machinery. ISBN 9798400...

2025
[15]

Data asset

National Institute of Standards and Technology . Data asset. https://csrc.nist.gov/glossary/term/data_asset, 2025. Accessed: June 27, 2025

2025
[16]

OpenAI . Chatgpt. https://chat.openai.com, 2025. Accessed: June 27, 2025

2025
[17]

Patil, Ion Stoica, and Joseph E

Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. Memgpt: Towards llms as operating systems, 2024. URL https://arxiv.org/abs/2310.08560

2024 arXiv
[18]

Collaborative memory: Multi-user memory sharing in llm agents with dynamic access control, 2025

Alireza Rezazadeh, Zichao Li, Ange Lou, Yuying Zhao, Wei Wei, and Yujia Bao. Collaborative memory: Multi-user memory sharing in llm agents with dynamic access control, 2025. URL https://arxiv.org/abs/2505.18279

2025 arXiv
[19]

Arnold, Chang Ye, Nick Masiewicki, Jinoo Baek, Peter Grabowski, and Eugene Ie

Nicholas Roth, Christopher Hidey, Lucas Spangher, William F. Arnold, Chang Ye, Nick Masiewicki, Jinoo Baek, Peter Grabowski, and Eugene Ie. Factored agents: Decoupling in-context learning and memorization for robust tool use, 2025. URL https://arxiv.org/abs/2503.22931

2025 arXiv
[20]

Agent workflow memory, 2024

Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. Agent workflow memory, 2024. URL https://arxiv.org/abs/2409.07429

2024 arXiv
[21]

Ai-native memory 2.0: Second me, 2025

Jiale Wei, Xiang Ying, Tao Gao, Fangyi Bao, Felix Tao, and Jingbo Shang. Ai-native memory 2.0: Second me, 2025. URL https://arxiv.org/abs/2503.08102

2025 arXiv
[22]

Autogen: Enabling next-gen llm applications via multi-agent conversation, 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. Autogen: Enabling next-gen llm applications via multi-agent conversation, 2023. URL https...

2023 arXiv
[23]

A-mem: Agentic memory for llm agents, 2025

Wujiang Xu, Kai Mei, Hang Gao, Juntao Tan, Zujie Liang, and Yongfeng Zhang. A-mem: Agentic memory for llm agents, 2025. URL https://arxiv.org/abs/2502.12110

2025 arXiv
[24]

Memorybank: Enhancing large language models with long-term memory, 2023

Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory, 2023. URL https://arxiv.org/abs/2305.10250

2023 arXiv
[25]

@esa (Ref

\@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...
[26]

\@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...
[27]

@open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Model context protocol

Anthropic . Model context protocol. GitHub repository, 2024. URL https://github.com/modelcontextprotocol. A protocol for connecting AI assistants to data sources and tools

work page 2024

[2] [2]

Anthropic . Claude. https://claude.ai, 2025. Accessed: June 27, 2025

work page 2025

[3] [3]

Mem0: Building production-ready ai agents with scalable long-term memory

Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:2504.19413, 2025

arXiv 2025

[4] [4]

Yiming Du, Wenyu Huang, Danna Zheng, Zhaowei Wang, Sebastien Montella, Mirella Lapata, Kam-Fai Wong, and Jeff Z. Pan. Rethinking memory in ai: Taxonomy, operations, topics, and future directions, 2025. URL https://arxiv.org/abs/2505.00675

arXiv 2025

[5] [5]

Memory sharing for large language model based agents, 2024

Hang Gao and Yongfeng Zhang. Memory sharing for large language model based agents, 2024. URL https://arxiv.org/abs/2404.09982

arXiv 2024

[6] [6]

Agent2agent (a2a) protocol

Google. Agent2agent (a2a) protocol. GitHub repository, 2024. URL https://github.com/a2aproject/A2A. Open-source framework for multi-agent systems

work page 2024

[7] [7]

Google . Gemini. https://gemini.google.com, 2025. Accessed: June 27, 2025

work page 2025

[8] [8]

Letters from future self: Augmenting the letter-exchange exercise with llm-based agents to enhance young adults' career exploration

Hayeon Jeon, Suhwoo Yoon, Keyeun Lee, Seo Hyeong Kim, Esther Hehsun Kim, Seonghye Cho, Yena Ko, Soeun Yang, Laura Dabbish, John Zimmerman, Eun-mee Kim, and Hajin Lim. Letters from future self: Augmenting the letter-exchange exercise with llm-based agents to enhance young adults' career exploration. In Proceedings of the 2025 CHI Conference on Human Factor...

arXiv 2025