The DOI matches its Crossref record; 39 of the 41 references it cites were found in a registry or an index; 42 of the 62 sources its claims cite were checked for support.
Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users
Open accessAbstract
Conversational AI is now the most widely adopted platform for interfacing with LLMs. Alongside LLMs these AI systems rely on contexts derived from past conversations and preferences to provide accurate and the most relevant responses to users. The knowledge base and past experiences contribute to long-term memory, while processing ongoing conversations generates short-term memory. Both long-term and short-term memories together provide a comprehensive and coherent context to the user. While most architectures focus on a single user context, there is an emerging need in conversational AI to provide a system to generate context from multiple individuals and/or agents. Building on this foundation, we introduce memory fabric, a framework that allows conversational AI to leverage context drawn from multiple users to generate coherent responses in a multiuser setting. This review is a synthesis of foundational memory architectures, from early models such as Neural Turing Machines and Key-Value Memory Networks to recent systems like AutoGen, SRMT, AgentVerse, and SAMEP evaluated through the lens of memory fabric. It also addresses work on dialogue and agent memory, multi-agent shared-memory designs, privacy and provenance strategies, benchmarking standards, and system-level considerations. A systematic review of English articles using IEEE Xplore, ACM, arXiv, SpringerLink, and ACL anthology databases was performed. Only full-text studies reported between 2014 and 2025 were included for analysis. Following PRISMA guidelines, a total of 662 studies were identified, of which 27 studies met the inclusion criteria. Overall, this review presents a detailed analysis on leveraging contexts from multiple users in a shared environment in foundational memory architectures, offering an in-depth analysis of their strengths and shortcomings, along with recommendations and future research directions.
Overview
CoChat Summary
TL;DR
This systematic review synthesizes memory-augmented architectures for conversational AI and introduces “memory fabric,” a framework for shared, persistent, multi-user and multi-agent memory. Following PRISMA, it screened 662 records and included 27 studies, then maps architectures, privacy/provenance, benchmarks, and open problems.
Key findings
- The PRISMA flow identified 662 studies (412 from databases, 250 from other sources), removed 342 duplicates, screened 320, reviewed 70 full texts, and included 27 studies in the detailed review.
- Foundational memory architectures such as Neural Turing Machines, Differentiable Neural Computers, Key-Value Memory Networks, and Dynamic Memory Networks provide addressable stores and adaptive read/write policies, but they assume a single logical owner; NTMs lacked access control and provenance, limiting direct multi-user use.
- The review categorizes multi-user shared conversational memory into tiered architectures, centralized shared memory with role-based policies, privacy-preserving shared fabrics, multi-agent memory ecosystems, and graph-structured memory with provenance.
- Benchmarks show persistent challenges: MADial-Bench finds updating long-term memories harder than retrieving them because models keep both current and outdated information; LoCoMo finds even LLMs with long context windows or retrieval augmentation struggle with coherence and continuity over long time frames, so human evaluation is still required.
Methods
- PRISMA-guided systematic review searching IEEE Xplore, ACM Digital Library, arXiv, SpringerLink, and ACL Anthology from January 2014 to May 2025, using Boolean search terms and inclusion/exclusion criteria; 27 studies met inclusion criteria.
- Data extraction covered architecture and memory type (NTM, DNC, SRMT), memory policy (episodic vs. persistent storage), privacy features such as access control and provenance, and benchmarking frameworks; authors independently extracted and evaluated data, resolving discrepancies through discussion.
Limitations
- The paper states limitations including privacy and security concerns, complexity in memory management, potential for inconsistent or conflicting memories, difficulty in evaluation, and scalability constraints.
- The paper also states unresolved gaps: absence of formal semantics for conflicting or overlapping memories, limited methods for evaluating privacy-utility tradeoffs, and lack of interaction paradigms to cultivate user trust, with governance, safety, and multi-party reliability layers remaining underdeveloped.
- Evident from the text: this is a review and synthesis that proposes memory fabric but reports no primary experiments, deployed system, or user study validating the framework empirically.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
28 Valid · 18 Not valid · 17 Not checked · 82 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“Benchmark for Memory-Augmented Dialogue Systems (MADial-Bench): MADial-Bench was developed to evaluate dialogue systems enhanced with external or persistent memory [22].”
- Also stated in: § 4 Benchmarking Methods · plain
“MADial-Bench was developed to evaluate dialogue systems enhanced with external or persistent memory [22].”
§ Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users > 4 Benchmarking Methods
What the sources say
- Supportedcites [22] 22. He, J., Li, W., Chen, K., Liu, W., Xu, J., Huang, M.: MADial-Bench: Towards Real-world Evaluation of Memory-Augmented Dialogue Generation. In: Pro- ceedings of the North American Chapter of the Association for Computational Linguistics (NAACL) (2025) source
The Introduction says MADial-Bench was introduced to evaluate memory-augmented dialogue systems and their generated responses. Section 3 describes the benchmark’s use of historical memories, matching the claim’s reference to persistent memory. (LLM, conf=0.94)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 4 Benchmarking Methods · plain
Every source the paper cites for this claim was checked and supports it.
In the paper
“Key-Value Memory Networks [13] further refined memory designs by separating key encoding for addressing and value encoding for output like a hash-map or dictionary.”
§ Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users > 3.1 Foundational Memory Architecture Review
What the sources say
- Supportedcites [13] 13. Miller, A.H., Fisch, A., Chen, J., Mott, A., Weston, J.: Key-Value Memory Net- works for Directly Reading Documents. In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1400-1409. Association for Computational Linguistics (2016) source
Section 3 says that lookup uses the key memory while reading the returned result uses the value memory. Section 3.1 describes memory slots as key–value pairs, supporting the hash-map or dictionary analogy. (LLM, conf=0.94)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Graves et al. (2014) introduced Neural Turing Machines (NTMs), which combined end-to-end differentiable controllers with an external memory matrix, enabling read and write operations via differentiable attention mechanisms.”
§ Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users > 3.1 Foundational Memory Architecture Review
What the sources say
- Supportedcites Graves et al. (2014) source
Section 3 (p. 5) describes an NTM as a neural network controller coupled to a memory matrix, with selective read and write operations governed by attentional weighting. The Abstract (p. 1) states that the combined system is differentiable end-to-end. (LLM, conf=0.95)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“The agents could plan future actions based on memories and reflective reasoning in a test environment.”
§ Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users > Memory Systems
What the sources say
- Supportedcites [16] 16. Park, J.S., O'Brien, W., Cai, H., Gao, M., Li, J., Narayanan, S., Lasko, A.: Generative Agents: Interactive Simulacra of Human Behavior. arXiv preprint arXiv:2304.03442 (2023) source
The Introduction describes an architecture that uses retrieved memories and reflections to form plans for future behavior. Section 6.1 says the agents were evaluated in a two-day Smallville simulation and interviewed about their ability to plan future actions based on their experiences. (LLM, conf=0.93)
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Benchmarks such as MADial-Bench provide practical evaluation protocols for memory recall and retrieval tasks, essential for measuring the correctness and privacy of multi-user memory systems [22].”
§ Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users > 3.2 Dialogue, Personalization, and Single-User Long-Term Memory
What the sources say
- Partly supportedcites [22] 22. He, J., Li, W., Chen, K., Liu, W., Xu, J., Huang, M.: MADial-Bench: Towards Real-world Evaluation of Memory-Augmented Dialogue Generation. In: Pro- ceedings of the North American Chapter of the Association for Computational Linguistics (NAACL) (2025) sourceExperimental — accuracy not yet measured
Sections 3.3 and 4.1.2 describe separate memory-recall and retrieval evaluations, including retrieval-accuracy metrics, supporting the claim that MADial-Bench provides an evaluation protocol for those tasks. The paper’s Introduction and Conclusion frame it as a benchmark for memory-augmented dialogue systems, but the cited passages do not establish a protocol for measuring privacy or for evaluating multi-user memory systems. (LLM, conf=0.90)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
30 more claims · 13 with their cited sources checked
Create a free account to see all 145 claims and to verify the 15 references still open.
- Probably not a substantial claim. 87
- It may state two separate claims as one. 61
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users > 3.4 Privacy, Protocols, and Governance for Shared Memories
“as noted in Masoor et al.'s SAMEP protocol (2025) [31].”
§ Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users > References
“31. Chen, K., Liu, M., Lee, S., Zhang, Y., Yu, Q.: SAMEP: Secure Agent Memory Exchange Protocol. arXiv preprint arXiv:2505.00911 (2025)”
Citations
Cites
Generative Agents: Interactive Simulacra of Human Behavior
Park, J. S., O'Brien, J., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023) · Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology · doi:10.1145/3586183.3606763 · Cited by 2,187
Generative Agents: Interactive Simulacra of Human Behavior
Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023) · arXiv · doi:10.48550/arxiv.2304.03442 · Cited by 2,173
Hybrid computing using a neural network with dynamic external memory
Graves, A., Wayne, G., Reynolds, M., Harley, T., Danihelka, I., Grabska-Barwińska, A., Colmenarejo, S. G., Grefenstette, E., Ramalho, T., Agapiou, J., et al. (2016) · Nature · doi:10.1038/nature20101 · Cited by 1,412
Conversational agents in healthcare: a systematic review
Laranjo, L., Dunn, A. G., Tong, H. L., Kocaballi, A. B., Chen, J., Bashir, R., Surian, D., Gallego, B., Magrabi, F., Lau, A. Y. S., et al. (2018) · Journal of the American Medical Informatics Association · doi:10.1093/jamia/ocy072 · Cited by 1,388
Personalizing Dialogue Agents: I have a dog, do you have pets too?
Zhang, S., Dinan, E., Urbanek, J., Szlam, A., Kiela, D., & Weston, J. (2018) · Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p18-1205 · Cited by 1,177
Graves, A., Wayne, G., & Danihelka, I. (2014) · arXiv · doi:10.48550/arxiv.1410.5401 · Cited by 1,058
(2016) · doi:10.18653/v1/d16-1147 · Cited by 1,035
Chatbot e-service and customer satisfaction regarding luxury brands
Chung, M., Ko, E., Joung, H., & Kim, S. J. (2020) · Journal of Business Research · doi:10.1016/j.jbusres.2018.10.004 · Cited by 998
The rise and potential of large language model based agents: a survey
Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al. (2025) · Science China Information Sciences · doi:10.1007/s11432-024-4222-0 · Cited by 970
A Persona-Based Neural Conversation Model
Li, J., Galley, M., Brockett, C., Spithourakis, G., Gao, J., & Dolan, B. (2016) · Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p16-1094 · Cited by 898
Research works printed but not linked: 1
- [14]14. Kumar, A., Irsoy, O., Ondruska, P., Iyyer, M., Bradbury, J., Gulrajani, I., Zhong, V., Paulus, R., Socher, R.: Ask Me Anything: Dynamic Memory Networks for Nat- ural Language Processing. In: Proceedings of the 33rd International Conference on Machine Learning, pp. 1378-1387. PMLR (2016)
Cited by
Adaptive Memory Retrieval Algorithms for Long-Term Context Retention in Autonomous Agents
Vairavan, D. S., Aswathy, D. S. U., Anitha, K., John, D. E. P., Rajan, D. C., & Sinha, A. (2026) · International Journal of Artificial Intelligence and Machine Learning · doi:10.51483/ijaiml.6.2s.2026.781-787
LLMs in Cloud: A Discussion on Architecture, Orchestration, and InferencePreprint
Noor, J., Rafee, A. N. M., Haque, M. E., Gupta, S. D., Osman, A., & Zishan, M. A. O. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.33475
Metadata
- Authors
- Tiwari, A., & Gupta, V.
- Affiliations
- Microsoft (United States) (US)
- The University of Texas Medical Branch at Galveston (US)
- Corresponding author
- Gupta, Vibhuti
- Venue
- TechRxiv
- Published
- 2025-12-08
- Volume
- 6
- Issue
- 1
- ISSN
- 2731-0809 (ISSN-L)
- OpenAlex
- W7129728849
- Cited by
- 2
- References
- 28
- Citations by year
- 2026: 2
- FWCI
- 2.97
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Discover Artificial Intelligence (publishedVersion, cc-by-nc-nd)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
- Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
- Social Sciences › Business, Management and Accounting › Marketing › AI in Service Interactions
- Keywords
- conversational AI agents, shared memory, long-term memory, short-term memory, context generation, neural Turing machines, AutoGen, agent memory, privacy, provenance, benchmarking
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users. https://app.cochat.ai/papers/10.36227/techrxiv.176523350.08289935/v1
- The paper
- Tiwari, A., & Gupta, V. (2025). Memory Fabric for Conversational AI Agents: Enabling Shared and Persistent Memory Across Users. TechRxiv. https://doi.org/10.36227/techrxiv.176523350.08289935/v1
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.