pith. sign in

arxiv: 2606.06240 · v1 · pith:FCSPMVVPnew · submitted 2026-06-04 · 💻 cs.DB · cs.AI

TOKI: A Bitemporal Operator Algebra for Contradiction Resolution in LLM-Agent Persistent Memory

Ziming Wang This is my paper

Pith reviewed 2026-06-27 22:59 UTC · model grok-4.3

classification 💻 cs.DB cs.AI
keywords bitemporal operatorscontradiction resolutionLLM agent memorywrite-time concurrency controlisolation levelsprovenanceaudit rowspersistent memory
0
0 comments X

The pith

TOKI types four memory heuristics as bitemporal operators that declare isolation levels and retain audit rows for LLM-agent writes.

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

The paper shows that contradiction resolution in persistent LLM-agent memory is write-time concurrency control, yet deployed heuristics never state the isolation level they assume or the anomalies they permit. TOKI supplies a single family of operators over a dual-row schema, each carrying an explicit isolation precondition and a provenance annotation that keeps the discarded fact in an audit row. Soundness theorems establish the guarantees for individual operators, pipelines, and n-ary conflicts. A tightness result proves that keyed logging of the language-model judge is required for replay consistency. A cross-system verdict matrix finds that every baseline keeping the judge admits at least one of replay inconsistency, belief-drift skew, or audit erasure, while TOKI excludes all three.

Core claim

TOKI types the four standard resolution heuristics as bitemporal operators over a dual-row schema. Each operator carries an isolation precondition and a provenance annotation that preserves the losing fact. Four soundness theorems close the contract across isolation, schema, and provenance, lift the guarantees to operator pipelines, and extend the fold operators to n-ary conflict sets. A tightness companion proves that keyed logging of the adjudicating judge is necessary for replay consistency. The verdict matrix shows TOKI alone excludes replay inconsistency, belief-drift skew, and audit erasure while retaining the language-model judge on the write path.

What carries the argument

The bitemporal operator algebra, which assigns each heuristic an isolation precondition and provenance annotation over a dual-row schema to preserve audit information while supporting operator composition.

If this is right

  • Each of the four heuristics now carries a declared isolation precondition and a provenance mechanism.
  • Losing facts remain accessible in audit rows instead of being erased.
  • Soundness extends from single operators to pipelines and to n-ary conflict sets.
  • Keyed logging of the language-model judge becomes a necessary condition for replay consistency.
  • Any system that removes the judge to avoid anomalies sacrifices the ability to perform judge-based adjudication.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same operator-typing discipline could be applied to read paths to obtain end-to-end consistency contracts.
  • Empirical validation on additional workloads beyond the single LoCoMo slice would test whether the relational schedule model holds in practice.
  • Content-addressed engines that avoid anomalies by dropping the judge may require separate mechanisms if nuanced, context-dependent resolution is later needed.
  • The dual-row schema plus provenance annotation supplies a concrete implementation target for production memory layers that currently use ad-hoc rules.

Load-bearing premise

The relational schedule model is an appropriate abstraction for modeling the write operations and conflict adjudication performed by LLM agents.

What would settle it

An execution trace on a TOKI implementation that exhibits replay inconsistency, belief-drift skew, or audit erasure on a workload whose schedule satisfies the stated isolation precondition would falsify the soundness theorems.

Figures

Figures reproduced from arXiv: 2606.06240 by Ziming Wang.

Figure 1
Figure 1. Figure 1: Contradiction resolution as write-time concurrency control. A bitemporal substrate detects a contradicting pair [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Each defence moves its primary slice while the con [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Contradiction-resolution latency stays flat as the [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Three structural anchors; every predicted [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The measured replay-inconsistency rate lands on [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: N1 admits across five frontier judges on the PARTIAL band. Disagreement rate by judge [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sub-linear scaling on four workload axes; 30 runs per point after 3 warmups, [PITH_FULL_IMAGE:figures/full_fig_p031_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The operator-to-isolation mapping is the Berenson–Adya anomaly lattice. Admit rate of each write-time anomaly [PITH_FULL_IMAGE:figures/full_fig_p034_9.png] view at source ↗
read the original abstract

Persistent memory for an LLM agent is a write-heavy substrate: every belief update is a versioned write, and a new claim may contradict a stored one. Production systems use four resolution heuristics (last-writer-wins, evidence-weighted merge, await-confirmation, per-rule policy), yet none declares the isolation level it assumes or the write-time anomalies it admits. We show that contradiction resolution is write-time concurrency control and make the missing contract explicit. TOKI types the four heuristics as one family of bitemporal operators over a dual-row schema, each with an isolation precondition and a provenance annotation that preserves the losing fact in an audit row. Four soundness theorems close the contract across isolation, schema, and provenance, lift the guarantees to operator pipelines, and extend the fold operators to n-ary conflict sets. A tightness companion proves that, within the relational schedule model, keyed logging of the adjudicating judge is necessary for replay consistency, which every audited baseline omits. A verdict matrix over eight systems localizes the gap: every baseline that keeps a language-model judge on the write path admits at least one of three write-time anomalies (replay inconsistency, belief-drift skew, audit erasure); a content-addressed engine-layer comparator avoids them only by removing the judge, and TOKI alone excludes all three while keeping it. On its one natural-workload slice the audit-row defence moves LoCoMo by 0.86, and ablating the typed memory layer removes 0.49 accuracy on 1,444 answerable LoCoMo questions; the cross-system comparison stays underpowered and claims no superiority. The contribution is the contract: a write-time correctness specification, proved sound across isolation, schema, and provenance, pinning the guarantee every production heuristic assumes but no deployed system makes explicit.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that contradiction resolution for LLM-agent persistent memory is a write-time concurrency control problem. It models four production heuristics as a family of bitemporal operators over a dual-row schema, each equipped with an isolation precondition and provenance annotation. Four soundness theorems are asserted to close the contract on isolation, schema, and provenance (including lifting to pipelines and n-ary conflicts), together with a tightness result proving keyed logging of the adjudicating judge is necessary for replay consistency inside the relational schedule model. A verdict matrix over eight systems shows that every baseline retaining an LM judge on the write path admits at least one of replay inconsistency, belief-drift skew, or audit erasure, while TOKI excludes all three; a single underpowered LoCoMo slice reports a 0.86 gain from audit rows and a 0.49 accuracy drop when the typed layer is ablated.

Significance. If the relational schedule model faithfully captures LLM-agent writes, the work supplies an explicit, proved correctness contract that current heuristics lack, together with a machine-checked-style localization of gaps via the verdict matrix and a tightness result on logging necessity. These formal elements would be a genuine contribution to the specification of persistent memory for agents. The empirical component is too narrow to support practical claims.

major comments (2)
  1. [relational schedule model and soundness theorems] The four soundness theorems and the tightness result on keyed logging are proved exclusively inside the relational schedule model with bitemporal operators and dual-row schema. The manuscript supplies neither a mapping from real LLM-agent write operations and semantic contradictions to these deterministic operators nor any empirical check that the model preserves the three anomalies it claims to exclude. This assumption is load-bearing for the central claim that TOKI alone excludes replay inconsistency, belief-drift skew, and audit erasure while retaining the LM judge.
  2. [verdict matrix] Verdict matrix (abstract): the localization that 'TOKI alone excludes all three' anomalies rests on the same unvalidated relational abstraction; without evidence that LLM writes satisfy the schedule assumptions, the matrix does not establish the claimed superiority outside the model.
minor comments (2)
  1. [abstract] Abstract: the four soundness theorems and tightness result are asserted without statements or proof sketches, forcing readers to reach the body to evaluate the formal contribution.
  2. [empirical evaluation] The cross-system comparison is explicitly labeled underpowered; this should be reflected more prominently when discussing the empirical slice.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting the foundational role of the relational schedule model. The contribution is a formal write-time correctness contract proved inside that model; we address the two major comments below and will revise the manuscript to clarify scope and add an explicit mapping discussion.

read point-by-point responses

  1. Referee: [relational schedule model and soundness theorems] The four soundness theorems and the tightness result on keyed logging are proved exclusively inside the relational schedule model with bitemporal operators and dual-row schema. The manuscript supplies neither a mapping from real LLM-agent write operations and semantic contradictions to these deterministic operators nor any empirical check that the model preserves the three anomalies it claims to exclude. This assumption is load-bearing for the central claim that TOKI alone excludes replay inconsistency, belief-drift skew, and audit erasure while retaining the LM judge.

    Authors: We agree the manuscript provides no explicit mapping from arbitrary real LLM writes to the deterministic operators and no empirical validation that the three anomalies appear in deployed systems. The relational schedule model is offered as a formal abstraction that isolates the write-time concurrency aspects of contradiction resolution, enabling the four soundness theorems and tightness result. The central claim is therefore scoped to properties that hold inside the model. We will add a new subsection (Section 3.5) that illustrates how common LLM-agent write patterns—belief updates from tool responses, user corrections, and multi-turn evidence accumulation—map onto sequences of the four typed operators, together with a limitations paragraph stating that empirical confirmation of anomaly occurrence in production traces remains future work. revision: partial

  2. Referee: [verdict matrix] Verdict matrix (abstract): the localization that 'TOKI alone excludes all three' anomalies rests on the same unvalidated relational abstraction; without evidence that LLM writes satisfy the schedule assumptions, the matrix does not establish the claimed superiority outside the model.

    Authors: The verdict matrix classifies each of the eight systems by whether it retains an LM judge on the write path and whether it satisfies the isolation, schema, and provenance preconditions of the relational schedule model. The localization therefore holds inside the model. We will revise the caption and surrounding text of the verdict matrix (Table 1) and the abstract to state explicitly that the comparison is performed under the assumptions of the relational schedule model and that claims of superiority are likewise scoped to that model. revision: partial

Circularity Check

0 steps flagged

No circularity; soundness theorems are model-internal and independent

full rationale

The paper introduces bitemporal operators over a dual-row schema, states four soundness theorems that close the contract across isolation/schema/provenance, and proves a tightness result inside the relational schedule model. These steps are definitional and deductive within the chosen abstraction; no equation, fitted parameter, or prediction is shown to reduce to its own inputs by construction. No self-citations appear as load-bearing premises, and the external applicability of the schedule model to LLM writes is an assumption rather than a circular derivation. The contribution is therefore self-contained against its stated model.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Review performed on abstract only; the ledger therefore records only the modeling assumptions explicitly invoked in the abstract.

axioms (1)
  • domain assumption The relational schedule model is the correct abstraction for LLM-agent write operations and judge adjudication.
    Invoked by the tightness companion proof that keyed logging is necessary for replay consistency.
invented entities (2)
  • Bitemporal operators no independent evidence
    purpose: Unified family that types the four heuristics with isolation preconditions and provenance annotations.
    Newly introduced construct whose properties are proved in the paper.
  • Dual-row schema no independent evidence
    purpose: Stores both winning and losing facts so that audit rows preserve provenance.
    Core schema design element of the proposed operators.

pith-pipeline@v0.9.1-grok · 5861 in / 1384 out tokens · 25787 ms · 2026-06-27T22:59:49.877261+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

87 extracted references · 27 canonical work pages · 3 internal anchors

  1. [1]

    Seyed Moein Abtahi, Rasa Rahnema, Hetkumar Patel, Neel Patel, Majid Fekri, and Tara Khani. 2026. Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents.arXiv preprint arXiv:2604.22085(2026).https://arxiv.org/abs/2604.22085

    Pith/arXiv arXiv 2026

  2. [2]

    Atul Adya, Barbara Liskov, and Patrick E. O’Neil. 2000. Generalized Isolation Level Definitions. InProceedings of the 16th International Conference on Data Engineering (ICDE). 67–78.https://doi.org/10.1109/ICDE.2000.839388

    work page doi:10.1109/icde.2000.839388 2000

  3. [3]

    James F. Allen. 1983. Maintaining Knowledge about Temporal Intervals.Commun. ACM26, 11 (1983), 832–843.https://doi.org/10.1145/182.358434

    work page doi:10.1145/182.358434 1983

  4. [4]

    Bahareh Sadat Arab, Su Feng Lee, Boris Glavic, Xing Niu, Seokki Lee, and Thomas Heinis. 2017. Using Reenactment to Retroactively Capture Provenance for Transactions. In Proceedings of the 2017 IEEE 33rd International Conference on Data Engineering (ICDE 2017). IEEE, 1077–1088.https://doi.org/10.1109/ICDE.2017.149

    work page doi:10.1109/icde.2017.149 2017

  5. [5]

    Pratyay Banerjee, Masud Moshtaghi, Shivashankar Subramanian, Amita Misra, and Ankit Chadha. 2026. APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI.arXiv preprint arXiv:2604.14362(2026).https://arxiv.org/abs/2604.14362

    Pith/arXiv arXiv 2026

  6. [6]

    Manuel Barros, Alcino Cunha, Jose Pereira, and Eunsuk Kang. 2026. Reasoning about Transactional Isolation Levels with Isolde.arXiv preprint arXiv:2604.00159(2026). https://arxiv.org/abs/2604.00159

    arXiv 2026

  7. [7]

    Bernstein, Jim Gray, Jim Melton, Elizabeth J

    Hal Berenson, Philip A. Bernstein, Jim Gray, Jim Melton, Elizabeth J. O’Neil, and Patrick E. O’Neil. 1995. A Critique of ANSI SQL Isolation Levels. InProceedings of the 1995 ACM SIGMOD International Conference on Management of Data. ACM, 1–10.https://doi.org/10.1145/223784.223785

    work page doi:10.1145/223784.223785 1995

  8. [8]

    Bernstein, Vassos Hadzilacos, and Nathan Goodman

    Philip A. Bernstein, Vassos Hadzilacos, and Nathan Goodman. 1987.Concurrency Control and Recovery in Database Systems. Addison-Wesley.https://www.microsoft.com/en- us/research/wp-content/uploads/2016/05/ccontrol.zip

    1987

  9. [9]

    Nishant Bhargava and Rodrigo Sobral Barrento. 2026. MemAudit: An Exact Package-Oracle Evaluation Protocol for Budgeted Long-Term LLM Memory Writing. arXiv:2605.02199 [cs.AI]https://arxiv.org/abs/2605.02199

    Pith/arXiv arXiv 2026

  10. [10]

    Richard Booth and Eva Richter. 2012. On Revising Fuzzy Belief Bases.arXiv preprint arXiv:1212.2444(2012).https://arxiv.org/abs/1212.2444

    Pith/arXiv arXiv 2012

  11. [11]

    Camille Bourgaux, Ana Ozaki, and Rafael Peñaloza. 2023. Semiring Provenance for Lightweight Description Logics.arXiv preprint arXiv:2310.16472(2023). https: //arxiv.org/abs/2310.16472

    arXiv 2023

  12. [12]

    Sophie Brinke, Anuj Dawar, Erich Grädel, and Benedikt Pago. 2026. Preservation Theorems in Semiring Semantics. arXiv:2605.10829 [cs.LO] https://arxiv.org/abs/2605. 10829

    Pith/arXiv arXiv 2026

  13. [13]

    Peter Buneman, Sanjeev Khanna, and Wang-Chiew Tan. 2001. Why and Where: A Characterization of Data Provenance. InDatabase Theory - ICDT 2001, 8th International Conference (Lecture Notes in Computer Science), Vol. 1973. Springer, 316–330.https://doi.org/10.1007/3-540-44503-X_20

    work page doi:10.1007/3-540-44503-x_20 2001

  14. [14]

    Cahill, Uwe Röhm, and Alan D

    Michael J. Cahill, Uwe Röhm, and Alan D. Fekete. 2008. Serializable Isolation for Snapshot Databases. InProceedings of the 2008 ACM SIGMOD International Conference on Management of Data. ACM, 729–738.https://doi.org/10.1145/1376616.1376690

    work page doi:10.1145/1376616.1376690 2008

  15. [15]

    Hanxiang Chao, Yihan Bai, Rui Sheng, Tianle Li, and Yushi Sun. 2026. STALE: Can LLM Agents Know When Their Memories Are No Longer Valid?arXiv preprint arXiv:2605.06527(2026).https://arxiv.org/abs/2605.06527

    Pith/arXiv arXiv 2026

  16. [16]

    Ding Chen, Simin Niu, Kehang Li, Peng Liu, Xiangping Zheng, Bo Tang, Xinchi Li, Feiyu Xiong, and Zhiyu Li. 2025. HaluMem: Evaluating Hallucinations in Memory Systems of Agents.arXiv preprint arXiv:2511.03506(2025).https://doi.org/10.48550/arXiv.2511.03506

    work page doi:10.48550/arxiv.2511.03506 2025

  17. [17]

    Ziyang Chen, Jinzhi Liao, and Xiang Zhao. 2023. Multi-granularity Temporal Question Answering over Knowledge Graphs. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 11378–11392. https://doi.org/10.18653/v1/2023.acl- long.637

    work page doi:10.18653/v1/2023.acl- 2023

  18. [18]

    Chris Cheney and Engram contributors. 2026. Engram: Bitemporal, Graph-Backed Memory System for AI Coding Agents. GitHub repository, AGPL-3.0 licence, latest release mcp@0.1.9(2025-12-30), HEAD commit 1553e53 (2026-01-05T20:10:48Z), 5 stars and 1 fork at survey.https://github.com/rawcontext/engram

    2026

  19. [19]

    James Cheney, Laura Chiticariu, and Wang-Chiew Tan. 2009. Provenance in Databases: Why, How, and Where.Foundations and Trends in Databases1, 4 (2009), 379–474. https://doi.org/10.1561/1900000006

    work page doi:10.1561/1900000006 2009

  20. [20]

    Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory.arXiv preprint arXiv:2504.19413(2025).https://arxiv.org/abs/2504.19413

    Pith/arXiv arXiv 2025

  21. [21]

    Cognitect. 2026. Datomic: An Information Model with Time, Provenance, and Accumulation. Online documentation. https://docs.datomic.com/datomic-overview.html

    2026

  22. [22]

    Pengfei Du. 2026. Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers.arXiv preprint arXiv:2603.07670(2026). https://arxiv.org/abs/ 2603.07670

    arXiv 2026

  23. [23]

    Alan Fekete, Dimitrios Liarokapis, Elizabeth O’Neil, Patrick O’Neil, and Dennis Shasha. 2005. Making Snapshot Isolation Serializable.ACM Transactions on Database Systems 30, 2 (2005), 492–528.https://doi.org/10.1145/1071610.1071615

    work page doi:10.1145/1071610.1071615 2005

  24. [24]

    Xpath evaluation in linear time

    J. Nathan Foster, Todd J. Green, and Val Tannen. 2008. Annotated XML: Queries and Provenance. InProceedings of the 27th ACM SIGMOD-SIGACT-SIGART Symposium on Principles of Database Systems (PODS ’08). ACM, 271–280.https://doi.org/10.1145/1376916.1376953

    work page doi:10.1145/1376916.1376953 2008

  25. [25]

    Harish Santhanalakshmi Ganesan. 2026. WorldDB: A Vector Graph-of-Worlds Memory Engine with Ontology-Aware Write-Time Reconciliation. arXiv:2604.18478 [cs.AI]. https://arxiv.org/abs/2604.18478

    Pith/arXiv arXiv 2026

  26. [26]

    Floris Geerts, Antonella Poggi, and Val Tannen. 2013. On the Limitations of Provenance for Queries with Difference. InProceedings of the Theory and Practice of Provenance Workshop (TaPP ’13). USENIX Association.https://doi.org/10.1145/2448496.2448516

    work page doi:10.1145/2448496.2448516 2013

  27. [27]

    Shabnam Ghasemirad, Si Liu, Christoph Sprenger, and David Basin. 2025. VerIso: Verifiable Isolation Guarantees for Database Transactions.Proc. VLDB Endow.18, 5 (2025), 1362–1375.https://doi.org/10.14778/3718057.3718065 41 Ziming Wang

    work page doi:10.14778/3718057.3718065 2025

  28. [28]

    Rohit Ghumare and AgentMemory contributors. 2026. AgentMemory: Persistent Memory for AI Coding Agents. GitHub repository, Apache-2.0 licence, npm package @agentmemory/agentmemory, latest release v0.9.17 (2026-05-16), HEAD commit c93c715 (2026-05-17T11:15:59Z).https://github.com/rohitg00/agentmemory

    2026

  29. [29]

    Erich Grädel and Val Tannen. 2024. Provenance Analysis and Semiring Semantics for First-Order Logic.arXiv preprint arXiv:2412.07986(2024). https://arxiv.org/abs/ 2412.07986

    arXiv 2024

  30. [30]

    Green, Grigoris Karvounarakis, and Val Tannen

    Todd J. Green, Grigoris Karvounarakis, and Val Tannen. 2007. Provenance Semirings. InProceedings of the 26th ACM SIGACT-SIGMOD-SIGART Symposium on Principles of Database Systems (PODS). 31–40.https://doi.org/10.1145/1265530.1265535

    work page doi:10.1145/1265530.1265535 2007

  31. [31]

    Green and Val Tannen

    Todd J. Green and Val Tannen. 2017. The Semiring Framework for Database Provenance. InProceedings of the 36th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems (PODS). 93–99.https://doi.org/10.1145/3034786.3056125

    work page doi:10.1145/3034786.3056125 2017

  32. [32]

    Bernal Jiménez Gutiérrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models. In Advances in Neural Information Processing Systems 37 (NeurIPS 2024).https://arxiv.org/abs/2405.14831

    arXiv 2024

  33. [33]

    Jun He and Deying Yu. 2026. Verifiable Agentic Infrastructure: Proof-Derived Authorization for Sovereign AI Systems. arXiv:2605.15228 [cs.AI] https://arxiv.org/abs/ 2605.15228

    Pith/arXiv arXiv 2026

  34. [34]

    Stratos Idreos. 2025. Alphabets, Grammars, Calculators, and the End of Hand-Crafted Systems.Proceedings of the VLDB Endowment (PVLDB)18, 12 (2025), 5537. https: //doi.org/10.14778/3750601.3760522

    work page doi:10.14778/3750601.3760522 2025

  35. [35]

    Jensen, Curtis E

    Christian S. Jensen, Curtis E. Dyreson, Michael Böhlen, James Clifford, Ramez Elmasri, Shashi K. Gadia, Fabio Grandi, Pat Hayes, Sushil Jajodia, et al. 1998. The Consensus Glossary of Temporal Database Concepts—February 1998 Version. InTemporal Databases: Research and Practice, Opher Etzion, Sushil Jajodia, and Suryanarayana Sripada (Eds.). Lecture Notes ...

    work page doi:10.1007/bfb0053710 1998

  36. [36]

    Sajjad Khan. 2026. S-Bus: Automatic Read-Set Reconstruction for Multi-Agent LLM State Coordination. arXiv:2605.17076 [cs.DC]https://arxiv.org/abs/2605.17076

    Pith/arXiv arXiv 2026

  37. [37]

    Krishna Kulkarni and Jan-Eike Michels. 2012. Temporal Features in SQL:2011.ACM SIGMOD Record41, 3 (2012), 34–43.https://doi.org/10.1145/2380776.2380786

    work page doi:10.1145/2380776.2380786 2012

  38. [38]

    Letta Team. 2026. Context Constitution: Letta Code’s Memory Filesystem. Letta engineering blog.https://letta.com/blog/context-constitution

    2026

  39. [39]

    Weixian Waylon Li, Jiaxin Zhang, Xianan Jim Yang, Tiejun Ma, and Yiwen Guo. 2026. Time is Not a Label: Continuous Phase Rotation for Temporal Knowledge Graphs and Agentic Memory.arXiv preprint arXiv:2604.11544(2026).https://arxiv.org/abs/2604.11544

    Pith/arXiv arXiv 2026

  40. [40]

    Zhiyu Li, Chenyang Xi, Chunyu Li, Ding Chen, Boyu Chen, Shichao Song, Simin Niu, Hanyu Wang, Jiawei Yang, Chen Tang, Qingchen Yu, Jihao Zhao, Yezhaohui Wang, Peng Liu, Zehao Lin, Pengyuan Wang, Jiahao Huo, Tianyi Chen, Kai Chen, Kehang Li, Zhen Tao, Huayi Lai, Hao Wu, Bo Tang, Zhengren Wang, Zhaoxin Fan, Ningyu Zhang, Linfeng Zhang, Junchi Yan, Mingchuan ...

  41. [41]

    MemOS: A Memory OS for AI System.arXiv preprint arXiv:2507.03724(2025).https://doi.org/10.48550/arXiv.2507.03724

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2507.03724 2025

  42. [42]

    Gonzalez, and Aditya G

    Shu Liu, Soujanya Ponnapalli, Shreya Shankar, Sepanta Zeighami, Alan Zhu, Shubham Agarwal, Ruiqi Chen, Samion Suwito, Shuo Yuan, Ion Stoica, Matei Zaharia, Alvin Cheung, Natacha Crooks, Joseph E. Gonzalez, and Aditya G. Parameswaran. 2026. Supporting Our AI Overlords: Redesigning Data Systems to be Agent-First. Proc. CIDR 2026, January 18–21, Chaminade.ht...

    2026

  43. [43]

    Lorentzos and Roger G

    Nikos A. Lorentzos and Roger G. Johnson. 1988. Extending Relational Algebra to Manipulate Temporal Data. InProceedings of the 14th International Conference on Very Large Data Bases (VLDB). Morgan Kaufmann, 289–296.https://vldb.org/conf/1988/P289.PDF

    1988

  44. [44]

    Yihao Lu, Wanru Cheng, Zeyu Zhang, and Hao Tang. 2026. MMA: Multimodal Memory Agent. arXiv:2602.16493 [cs.AI]https://arxiv.org/abs/2602.16493

    arXiv 2026

  45. [45]

    Yang Luo, Zifeng Kang, Tiantian Ji, Xinran Liu, Yong Liu, Shuyu Li, and Lingyun Peng. 2026. ShadowMerge: A Novel Poisoning Attack on Graph-Based Agent Memory via Relation-Channel Conflicts.arXiv preprint arXiv:2605.09033(2026).https://arxiv.org/abs/2605.09033

    Pith/arXiv arXiv 2026

  46. [46]

    Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. 2024. Evaluating Very Long-Term Conversational Memory of LLM Agents. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 13851–13870. https://doi.org/10.18653/v1/2024.acl- long.747

    work page doi:10.18653/v1/2024.acl- 2024

  47. [47]

    Mem0 Team. 2026. Mem0 README: New Memory Algorithm (April 2026). GitHub repository README.https://github.com/mem0ai/mem0/blob/main/README.md

    2026

  48. [48]

    Hua Meng, Zhiguo Long, Michael Sioutis, and Zhengchun Zhou. 2025. On Definite Iterated Belief Revision with Belief Algebras.arXiv preprint arXiv:2505.06505(2025). https://arxiv.org/abs/2505.06505

    arXiv 2025

  49. [49]

    Solomon Messing. 2026. Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking. arXiv:2604.11581 [cs.CL] https://arxiv.org/ abs/2604.11581

    Pith/arXiv arXiv 2026

  50. [50]

    Bardia Mohammadi, Nearchos Potamitis, Lars Klein, Akhil Arora, and Laurent Bindschaedler. 2026. Atomix: Timely, Transactional Tool Use for Reliable Agentic Workflows. arXiv:2602.14849 [cs.DC]https://arxiv.org/abs/2602.14849

    Pith/arXiv arXiv 2026

  51. [51]

    Mehryar Mohri. 2002. Semiring Frameworks and Algorithms for Shortest-Distance Problems.Journal of Automata, Languages and Combinatorics7, 3 (2002), 321–350. http://www.cs.nyu.edu/~mohri/pub/jalc.pdf

    2002

  52. [52]

    Praveen Kumar Myakala, Manan Agrawal, and Rahul Manche. 2026. BeliefShift: Benchmarking Temporal Belief Consistency and Opinion Drift in LLM Agents.arXiv preprint arXiv:2603.23848(2026).https://arxiv.org/abs/2603.23848

    arXiv 2026

  53. [53]

    Ciyan Ouyang and Rui Hou. 2026. MemLineage: Lineage-Guided Enforcement for LLM Agent Memory.arXiv preprint arXiv:2605.14421(2026). https://arxiv.org/abs/ 2605.14421

    Pith/arXiv arXiv 2026

  54. [54]

    Patil, Ion Stoica, and Joseph E

    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).https://arxiv.org/abs/2310.08560

    Pith/arXiv arXiv 2023

  55. [55]

    O’Brien, Carrie J

    Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST). 1–22.https://doi.org/10.1145/3586183.3606763

    work page doi:10.1145/3586183.3606763 2023

  56. [56]

    Young Bin Park. 2026. Graph-Native Cognitive Memory for AI Agents: Formal Belief Revision Semantics for Versioned Memory Architectures.arXiv preprint arXiv:2603.17244 (2026).https://arxiv.org/abs/2603.17244

    arXiv 2026

  57. [57]

    Liana Patel, Siddharth Jha, Melissa Pan, Harshit Gupta, Parth Asawa, Carlos Guestrin, and Matei Zaharia. 2025. Semantic Operators and Their Optimization: Enabling LLM-Based Data Processing with Accuracy Guarantees in LOTUS.Proceedings of the VLDB Endowment (PVLDB)18, 11 (2025), 4171–4184. https://doi.org/10.14778/3749646.3749685

    work page doi:10.14778/3749646.3749685 2025

  58. [58]

    Sidharth Pulipaka, Stanislau Hlebik, Leonidas Raghav, Sahar Abdelnabi, Vyas Raina, Ivaxi Sheth, and Mario Fritz. 2026. Hidden in Memory: Sleeper Memory Poisoning in LLM Agents. arXiv:2605.15338 [cs.CR]https://arxiv.org/abs/2605.15338

    Pith/arXiv arXiv 2026

  59. [59]

    Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. 2025. Zep: A Temporal Knowledge Graph Architecture for Agent Memory.arXiv preprint arXiv:2501.13956(2025).https://arxiv.org/abs/2501.13956

    Pith/arXiv arXiv 2025

  60. [60]

    Michaël Roynard. 2026. The Missing Knowledge Layer in Cognitive Architectures for AI Agents.arXiv preprint arXiv:2604.11364(2026). https://arxiv.org/abs/2604.11364

    Pith/arXiv arXiv 2026

  61. [61]

    Matthew Russo, Chunwei Liu, Sivaprasad Sudhir, Gerardo Vitagliano, Michael Cafarella, Tim Kraska, and Samuel Madden. 2026. Abacus: A Cost-Based Optimizer for Semantic Operator Systems.Proceedings of the VLDB Endowment (PVLDB)19, 5 (2026), 1060–1073.https://www.vldb.org/pvldb/vol19/p1060-russo.pdf

    2026

  62. [62]

    Rashomon Memory: Towards Argumentation-Driven Retrieval for Multi-Perspective Agent Memory

    Albert Sadowski and Jarosław A. Chudziak. 2026. Rashomon Memory: Towards Argumentation-Driven Retrieval for Multi-Perspective Agent Memory. https://doi.org/ 10.48550/arXiv.2604.03588arXiv:2604.03588

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.03588arxiv:2604.03588 2026

  63. [63]

    Kenneth Salem, Hector Garcia-Molina, and Jeannie Sands. 1989. Altruistic Locking: A Strategy for Coping with Long Lived Transactions. InProceedings of the 2nd International Workshop on High Performance Transaction Systems.https://doi.org/10.1145/64162.64173

    work page doi:10.1145/64162.64173 1989

  64. [64]

    Dhravya Shah and Supermemory contributors. 2026. Supermemory: State-of-the-art Memory and Context Engine for AI. GitHub repository supermemoryai/supermemory, MIT licence, HEAD commit 36ecf47 (2026-05-17T07:55:12Z), 22,596 stars, zero GitHub releases, Python SDK supermemory-openai-sdk on PyPI at alpha v1.0.3. https: //github.com/supermemoryai/supermemory

    2026

  65. [65]

    Parameswaran, and Eugene Wu

    Shreya Shankar, Tristan Chambers, Tarak Shah, Aditya G. Parameswaran, and Eugene Wu. 2025. DocETL: Agentic Query Rewriting and Evaluation for Complex Document Processing.Proceedings of the VLDB Endowment (PVLDB)18, 12 (2025), 3920–3932.https://www.vldb.org/pvldb/vol18/p3920-shankar.pdf

    2025

  66. [66]

    Richard Snodgrass and Ilsoo Ahn. 1986. Temporal Databases.IEEE Computer19, 9 (1986), 35–42.https://doi.org/10.1109/MC.1986.1663327

    work page doi:10.1109/mc.1986.1663327 1986

  67. [67]

    Snodgrass

    Richard T. Snodgrass. 2000.Developing Time-Oriented Database Applications in SQL. Morgan Kaufmann, San Francisco, CA. https://www2.cs.arizona.edu/~rts/tdbbook. pdf 42 TOKI: A Bitemporal Operator Algebra for Contradiction Resolution in LLM-Agent Persistent Memory

    2000

  68. [68]

    Miao Su, Yucan Guo, Zhongni Hou, Long Bai, Zixuan Li, Yufei Zhang, Guojun Yin, Wei Lin, Xiaolong Jin, Jiafeng Guo, and Xueqi Cheng. 2026. Beyond Dialogue Time: Temporal Semantic Memory for Personalized LLM Agents.arXiv preprint arXiv:2601.07468(2026).https://arxiv.org/abs/2601.07468

    arXiv 2026

  69. [69]

    Tencent. 2026. TencentDB Agent Memory: Fully Local Long-Term Memory for AI Agents via a 4-Tier Progressive Pipeline. GitHub repository Tencent/TencentDB-Agent-Memory, MIT licence (Tencent header wrapper), HEAD commit 5736acc (2026-05-16T12:17:22Z), latest release v0.3.4 (2026-05-13), 2,674 stars, npm package@tencentdb-agent-memory/memory-tencentdb, Node >...

    2026

  70. [70]

    The XTDB Authors. 2026. XTDB 2.x: Bitemporal SQL for the Real World. Online documentation.https://docs.xtdb.com/intro/what-is-xtdb.html

    2026

  71. [71]

    Md Nayem Uddin, Kumar Shubham, Eduardo Blanco, Chitta Baral, and Gengyu Wang. 2026. From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents.arXiv preprint arXiv:2604.20006(2026).https://arxiv.org/abs/2604.20006

    Pith/arXiv arXiv 2026

  72. [72]

    Pruthvinath Jeripity Venkata. 2026. Three Regimes of Context-Parametric Conflict: A Predictive Framework and Empirical Validation. arXiv:2605.11574 [cs.CL] https: //arxiv.org/abs/2605.11574

    Pith/arXiv arXiv 2026

  73. [73]

    Joel Ward. 2025. MemoriesDB: A Temporal-Semantic-Relational Database for Long-Term Agent Memory.arXiv preprint arXiv:2511.06179(2025). https://arxiv.org/abs/ 2511.06179

    arXiv 2025

  74. [74]

    Lei Wei, Xiao Peng, Xu Dong, Niantao Xie, and Bin Wang. 2026. FadeMem: Biologically-Inspired Forgetting for Efficient Agent Memory.arXiv preprint arXiv:2601.18642(2026). https://arxiv.org/abs/2601.18642

    arXiv 2026

  75. [75]

    Albert Widiaatmaja, Belkis Djeffal, Ashish Dandekar, and Pierre Senellart. 2025. Demonstration of ProvSQL Update Provenance through Temporal Databases. InProvenance Week@SIGMOD.https://doi.org/10.1145/3736229.3736253

    work page doi:10.1145/3736229.3736253 2025

  76. [76]

    Di Wu, Zixiang Ji, Asmi Kawatkar, Bryan Kwan, Jia-Chen Gu, Nanyun Peng, and Kai-Wei Chang. 2026. LongMemEval-V2: Evaluating Long-Term Agent Memory Toward Experienced Colleagues.arXiv preprint arXiv:2605.12493(2026).https://arxiv.org/abs/2605.12493

    Pith/arXiv arXiv 2026

  77. [77]

    Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. 2025. LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. In Proceedings of the 13th International Conference on Learning Representations (ICLR).https://openreview.net/forum?id=pZiyCaVuti

    2025

  78. [78]

    Shiyao Xie and Jian Du. 2026. Neuro-Symbolic Resolution of Recommendation Conflicts in Multimorbidity Clinical Guidelines. InProceedings of the 40th Annual AAAI Conference on Artificial Intelligence, Bridge Program on Logic and AI.https://doi.org/10.48550/arXiv.2604.17340

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.17340 2026

  79. [79]

    Jingbo Yang, Kwei-Herng Lai, Xiaowen Wang, Shiyu Chang, Yaar Harari, and Evgeniy Gabrilovich. 2026. GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations. arXiv:2605.14498 [cs.CL]https://arxiv.org/abs/2605.14498

    Pith/arXiv arXiv 2026

  80. [80]

    Jiawei Yu, Yixiang Fang, Xilin Liu, and Yuchi Ma. 2026. H-Mem: A Novel Memory Mechanism for Evolving and Retrieving Agent Memory via a Hybrid Structure. arXiv:2605.15701 [cs.AI]https://arxiv.org/abs/2605.15701

    Pith/arXiv arXiv 2026

Showing first 80 references.