CoChat
Sign Up
Cited by 4AI-based Problem Solving and PlanningTopic ModelingNatural Language Processing Techniques

MemEngine: A Unified and Modular Library for Developing Advanced Memory of LLM-based Agents

Open access
Zhang, Zeyu, Dai, Quanyu, Chen, Xu, Li, Rui, Li, Zhongyang, Dong, Zhenhua·2025·Companion Proceedings of the ACM on Web Conference 2025
DOI 10.1145/3701716.3715299
Open this paper in CoChat — free accountOpen PDF

Abstract

Recently, large language model based (LLM-based) agents have been widely applied across various fields.As a critical part, their memory capabilities have captured significant interest from both industrial and academic communities.Despite the proposal of many advanced memory models in recent research, however, there remains a lack of unified implementations under a general framework.To address this issue, we develop a unified and modular library for developing advanced memory models of LLM-based agents, called MemEngine.Based on our framework, we implement abundant memory models from recent research works.Additionally, our library facilitates convenient and extensible memory development, and offers user-friendly and pluggable memory usage.For benefiting our community, we have made our project publicly available at https://github.com/nuster1128/MemEngine.

AI snapshot

Written from the paper’s full text.

TL;DR

MemEngine is presented as a unified, modular library for building and using memory components in LLM-based agents. It organizes memory into functions, operations, and models, and packages multiple published memory models with configuration, deployment, and customization support.

Key findings

  • Introduces a three-level framework: memory functions (e.g., encoder, retrieval, reflector, summarizer, trigger, utilization, forget, truncation, judge, LLM), memory operations (store, recall, manage, optimize), and memory models.
  • Implements nine named memory models: FUMemory, LTMemory, STMemory, GAMemory, MBMemory, SCMemory, MGMemory, RFMemory, and MTMemory, switchable through common interfaces such as reset, store, recall, manage, and optimize.
  • Provides configuration for hyperparameters and prompts, utilities for storage, display, remote client use, and automatic model selection, plus local/remote deployment, default/configurable/automatic usage modes, and compatibility with tools such as vllm and AutoGPT.
  • A qualitative comparison table indicates that most surveyed agent libraries and independent memory libraries offer plug-and-play integration and basic read/write support, while few support reflection and optimization; MemEngine claims comprehensive default models and advanced customization. No quantitative metrics are reported.

Methods

  • This is a system/library design paper: it proposes an architecture and implementation rather than an empirical study.
  • It describes modules, interfaces, usage workflows, and customization paths, and compares feature support qualitatively in Table 1.
  • Data or sample and statistical analysis: Not stated in the paper.

Limitations

  • The paper does not explicitly list limitations. Evident from the text: no quantitative evaluation, benchmark, or user study is reported, and the comparison is a qualitative feature table.
  • The paper mentions future work on multimodal memory, indicating that the current library focuses on textual memory.
  • The text notes that minor adjustments were made to fit models into the framework, with details deferred to documentation and source code, which limits what can be assessed from the paper alone.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3701716.3715299

Cites

16 of 16 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

7 of 7 cited sources checked · 6 supported · 0 flagged · accuracy not yet measured

  • Findingreasoning

    Based on its unified and modular framework, MemEngine implements a wide range of memory models from recent research works.

    “Based on our unified and modular framework, we implement a wide range of memory models from recent research works, many of which are widely applied in diverse applications.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The reasonable adaptations made when implementing the research memory models do not materially alter the behavior of the original models.
  • Findingnone given

    MemEngine provides a comprehensive set of default hyper-parameters and prompts, allowing developers and researchers to adjust only specific parts without altering others.

    “Second, we provide a comprehensive set of default hyper-parameters and prompts, where developers and researchers can adjust only the specific parts without altering others.”
    Quote in the paperSubstantial claim
  • Interpretationreasoning

    MemEngine is a unified and modular library that facilitates the development of advanced memory models for LLM-based agents.

    “In order to address the above problems, we develop a unified and modular library named MemEngine, which facilitates the development of advanced memory models for LLM-based agents.”
    • Also stated in: § Abstract · plain
      “To address this issue, we develop a unified and modular library for developing advanced memory models of LLM-based agents, called MemEngine.”
    Quote in the paperSubstantial claim
  • Interpretationreasoning

    MemEngine is the first library that implements a wide variety of memory models from research works under a unified and modular framework.

    “In summary, MemEngine is the first library that implements a wide variety of memory models from research works under a unified and modular framework, facilitating both convenient development and ease of use.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The reasonable adaptations made when implementing the research memory models do not materially alter the behavior of the original models.
  • Classified as probably not a substantial claim. 25
  • May combine two claims. 10
  • The quote only partly states this claim. 1

  • Probably not a substantial claim: may not pass S2 (checkable).

    Premisecitation

    Memory is one of the most critical components of LLM-based agents, determining how they store historical data, reflect on existing knowledge, and recall useful information to support decision-making.

    “Among various internal modules, memory is one of the most critical components for agents, as it determines how they store historical data, reflect on existing knowledge, and recall useful information to support decision-making [13].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [13] 13. Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2024. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501 (2024). source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    RFMemory (Reflexion) is a memory model that can learn to memorize from previous trajectories by optimization.

    “• RFMemory (Reflexion [9]): A prominent memory model that can learn to memorize from previous trajectories by optimization.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [9] 9. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2024. Reflexion: Language agents with verbal reinforcement learning. NeurIPS 36 (2024). source

      Section 3 describes an iterative optimization process in which the agent generates a trajectory, turns feedback on it into a self-reflection, and appends that reflection to memory. Section 5 calls Reflexion a natural-language policy-optimization technique. The paper supports learning from previous trajectories through memory, but it describes storing reflections rather than learning to memorize trajectories by optimizing the memory itself. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    MGMemory (MemGPT) is a hierarchical memory model that treats the memory system as an operating system.

    “• MGMemory (MemGPT [6]): A hierarchical memory model that treats the memory system as an operation system.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [6] 6. Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G Patil, Ion Stoica, and Joseph E Gonzalez. 2023. Memgpt: Towards llms as operating systems. arXiv preprint arXiv:2310.08560 (2023). source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    SCMemory (SCM) is a self-controlled memory model that can recall minimal but necessary information for inference.

    “• SCMemory (SCM [10]): A self-controlled memory model that can recall minimal but necessary information for inference.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [10] 10. Bing Wang, Xinnian Liang, Jian Yang, Hui Huang, Shuangzhi Wu, Peihao Wu, Lu Lu, Zejun Ma, and Zhoujun Li. 2023. Enhancing large language model with self-controlled memory framework. arXiv preprint arXiv:2304.13343 (2023). source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    MTMemory (MemTree) is a dynamic memory model with a tree-structured semantic representation to organize information.

    “• MTMemory (MemTree [8]): A dynamic memory model with a tree-structured semantic representation to organize information.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [8] 8. Alireza Rezazadeh, Zichao Li, Wei Wei, and Yujia Bao. 2024. From Isolated Conversations to Hierarchical Schemas: Dynamic Tree Memory Representation for LLMs. arXiv preprint arXiv:2410.14052 (2024). source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    MBMemory (MemoryBank) is a multi-layered memory model with dynamic summarization and a forgetting mechanism.

    “• MBMemory (MemoryBank [14]): A multi-layered memory model with dynamic summarization and forgetting mechanism.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [14] 14. Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Mem- orybank: Enhancing large language models with long-term memory. In AAAI, Vol. 38. 19724-19731. source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    GAMemory (Generative Agents) is a pioneer memory model with weighted retrieval combination and a self-reflection mechanism.

    “• GAMemory (Generative Agents [7]): A pioneer memory model with weighted retrieval combination and self-reflection mechanism.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [7] 7. Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In UIST. 1-22. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    MemEngine provides detailed instructions and examples in its documentation to guide customization by researchers.

    “To better support researchers' development, we provide detailed instructions and examples in our document to guide the customization.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Findingnone given

    MemEngine offers three modes for utilizing pre-implemented memory models: default, configurable, and automatic.

    “After deployment, there are three modes available for utilizing pre-implemented memory models.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    MemEngine implements its memory models with the interfaces reset, store, recall, manage, and optimize.

    “Specifically, these models are implemented with the interfaces including reset, store, recall, manage, and optimize.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    GAMemoryStore utilizes LLMJudge to provide the importance score on each observation.

    “For example, GAMemoryStore utilizes LLMJudge to provide the importance score on each observation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    The MemEngine project has been made publicly available at https://github.com/nuster1128/MemEngine.

    “For benefiting our community, we have made our project publicly available at https://github.com/nuster1128/MemEngine.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    MemEngine is compatible with several well-known tools and frameworks of LLM-based agents, such as vllm and AutoGPT.

    “Additionally, our library offers compatibility with several wellknown tools and frameworks, such as vllm and AutoGPT.”
    • Also stated in: § Convenient and Extensible Memory Development. · plain
      “Our library is also compatible with some prominent frameworks of LLM-based agents, such as AutoGPT.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    MemEngine's configuration module is hierarchical and corresponds to its three-level memory framework, enabling adjustment of both hyper-parameters and prompts within the memory models.

    “First, we design a hierarchical memory configuration module corresponding to our three-level memory framework, enabling adjustments to both hyper-parameters and prompts within the memory models.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    In the automatic mode, MemEngine automatically selects the appropriate memory models, hyper-parameters, and prompts from the provided ranges, based on a specific task's criteria.

    “In the automatic mode, the library automatically selects the appropriate memory models, hyper-parameters, and prompts from the provided ranges, based on a specific task's criteria.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    MemEngine implements a storage module as a database to retain the contents of information.

    “We implement a storage module as a database to retain the contents of information.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    MemEngine's configuration supports both statistic manners (e.g., files) and dynamic manners (e.g., dictionaries).

    “Finally, our configuration supports both statistic manners (e.g., files) and dynamic manners (e.g., dictionaries).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingnone given

    MTMemory and LTMemory share the common memory recall operation LTMemoryRecall, while MTMemory has its own memory store operation MTMemoryStore to implement tree-structured information update.

    “For example, MTMemory and LTMemory share the common memory recall operation LTMemoryRecall, while MTMemory has its own memory store operation MTMemoryStore to implement the tree-structured information update.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    Most existing libraries offer plug-and-play memory components to store and recall information for LLM-based agents, but few of them implement advanced memory operations like reflection and optimization.

    “While most of these libraries offer plug-and-play memory components to store and recall information for LLM-based agents, few of them implement advanced memory operations like reflection and optimization.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    MemEngine offers a client module to utilize MemEngine through remote deployment on a server implemented by FastAPI.

    “Besides, we offer a client module to utilize MemEngine through remote deployment on a server implemented by FastAPI.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    MemEngine provides a display module to visualize the specific contents within the memory.

    “We also provide a display module to visualize the specific contents within the memory.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    MemEngine implements an automatic selector to assist developers in choosing memory models with hyper-parameters for their own tasks.

    “We also implement an automatic selector to assist developers to choose memory models with hyper-parameters for their own tasks.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    MemEngine's memory modules are pluggable and can be easily utilized across different agent frameworks.

    “Moreover, our memory modules are pluggable and can be easily utilized across different agent frameworks.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Compared with other libraries, MemEngine implements comprehensive research models under a unified framework and provides modular operations and functions that assist researchers in customizing advanced memory models.

    “Moreover, compared with other libraries, as a major contribution, MemEngine implements comprehensive research models under a unified framework, along with providing modular operations and functions that assist researchers in customizing advanced memory models.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The reasonable adaptations made when implementing the research memory models do not materially alter the behavior of the original models.

  • May combine two claims.

    Interpretationreasoning

    All three levels of MemEngine's framework are modularized, with higher-level modules able to reuse lower-level modules, improving efficiency and consistency in implementation.

    “All these three levels are modularized inside our framework, where higher-level modules can reuse lower-level modules, thereby improving efficiency and consistency in implementation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Recent works have proposed various memory models for LLM-based agents, but these models are implemented under different pipelines and lack a unified framework.

    “Although some recent works have proposed various memory models for LLM-based agents, they are implemented under different pipelines and lack a unified framework.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    MemEngine's framework is composed of three hierarchical levels: memory functions at the lowest level, memory operations at the intermediate level, and memory models at the highest level, organized to implement existing research models under a general structure.

    “We propose a unified memory framework composed of three hierarchical levels to organize and implement existing research models under a general structure.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Many academic memory models are tightly integrated with agents in a non-pluggable manner, making them difficult to apply across different agents.

    “Besides, many academic models are tightly integrated with agents in a non-pluggable manner, making them difficult to apply across different agents.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Many basic memory functions, such as retrieval and summarization, are duplicated across different memory models, so researchers often need to implement them repeatedly when developing new models.

    “Moreover, many basic functions (such as retrieval and summarization) are duplicated across different models, and researchers often need to implement them repeatedly when developing new models.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Based on MemEngine's modular memory operations and memory functions, researchers can conveniently develop their own advanced memory models.

    “Based on our modular memory operations and memory functions, researchers can conveniently develop their own advanced memory models.”
    Quote in the paperSubstantial claim
30 of 30

  • The paper relies on this without support.

    The reasonable adaptations made when implementing the research memory models do not materially alter the behavior of the original models.

    “All of these memory models are implemented by combining various memory operations, and we make some reasonable adaptations”
1 of 1

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Multi-domain agents engage with tools and/or APIs, manage complex workflows such as coding, healthcare, finance, and crime analysis, and provide memory-based learning that allows them to continue a task by absorbing adaptive information from external cues.

    “Multi-domain agents engage with tools and/or APIs, manage workflows that will inevitably be very complex, such as coding, healthcare, finance, crime analysis, etc., and provide memory-based learning that allows them to continue a task by absorbing adaptive information from external cues [31,67,75-77,136,139].”

    From The Rise of Agentic AI: A Review of Definitions, Frameworks, Architectures, Applications, Evaluation Metrics, and Challenges (2025) · cites this paper as [139]

    Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

    Experimental — accuracy not yet measured

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Zhang, Z., Dai, Q., Chen, X., Li, R., Li, Z., & Dong, Z.
Affiliations
  • Renmin University of China (CN)
  • Huawei Technologies (China) (CN)
ORCID
Venue
Companion Proceedings of the ACM on Web Conference 2025
Publisher
ACM
Published
2025-05-08
Pages
821-824
DOI
10.1145/3701716.3715299
OpenAlex
W4410637010
Cited by
2
References
4
Citations by year
2026: 1 · 2025: 1
FWCI
1.90
Type
conference paper
Language
en
Open access
Licence
www.acm.org/publications/policies/copyright_policy#Background (vor)
Locations
  • International World Wide Web Conference (WWW) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National Natural Science Foundation of China
  • Renmin University of China
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
Keywords
LLM agents, memory models, modular framework, large language models, agent memory

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-11.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-11.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review