Elamur Research⌕
☰
home / papers / 2026 / Beyond Retrieval: Bi-Temporal State Arbitration for Longitudinal Healthcare Agents

Beyond Retrieval: Bi-Temporal State Arbitration for Longitudinal Healthcare Agents

Paper · 2026 · Jianing Zhao, Xiaoquan Zhi and Xinqiang Yu

Summary

LLM agents that follow patients with chronic conditions over months must know the patient's current state, such as medications, symptoms and vitals, even though these facts keep changing. The authors argue that existing memory designs fail at this because they are retrieval-centric. Vector RAG, Graphiti/Zep-style temporal graphs and long-context models return past fragments but do not settle which fact is still valid. The authors instead propose state-centric memory, drawing on Truth Maintenance Systems and bi-temporal databases.

Each patient attribute is stored as a CanonicalMemory unit. This unit separates when an event happened from when the agent learned about it, and it records a validity window. New evidence is merged incrementally through four arbitration operators, so conflicting reports are kept rather than overwritten. A confidence-thresholded escalation layer decides which memory tier answers each query.

On a longitudinal Type 2 diabetes benchmark, the system clearly outperforms long-context and vector-memory baselines. This suggests that long-horizon clinical agents need explicit arbitration, not just better retrieval.

Key points

  • Results: The method reaches Unique-F1 0.85 and Conflict-F1 0.98. Long-context LLMs score 0.38 / 0.89, and standard vector memory scores 0.30 / 0.60.
  • Arbitration operators:
    • SUPPORT raises confidence when a report repeats a known value.
    • REFINE merges added detail into the existing state.
    • SUPERSEDE closes the old validity window and opens a new state. It is triggered by lexical cues like "switched to".
    • BRANCH-CONFLICT keeps competing candidates side by side.
  • Confidence scoring: Each candidate value gets a score that mixes three factors:
    • source authority (clinician 1.0, patient self-report 0.5, inferred 0.3), weighted α=0.5;
    • recency, weighted β=0.3;
    • log-scaled count of corroborating mentions, weighted γ=0.2.
  • Error analysis:
    • Vague time expressions caused 22% of errors.
    • Overly strict semantic matching caused 12%, for example treating Glucophage vs. Metformin as a conflict.
  • Released resources: The evaluation framework and the Med-LongMem diabetes suite are released. The authors note that only one disease domain was tested and that SUPERSEDE detection depends on hand-written patterns.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Tianjin UniversityTianjin · China52 works · 1,128 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksMedicine · 275 worksHealth Informatics · 214 works

🧭 Topics 1Research subjects and languages the work is about.Information retrieval & search · 552 works

Cite

Cite as Jianing Zhao, Xiaoquan Zhi and Xinqiang Yu. 2026. Beyond Retrieval: Bi-Temporal State Arbitration for Longitudinal Healthcare Agents. Proceedings of the 4th Workshop on Towards Knowledgeable Foundation Models (KnowFM 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.knowfm-1.10

@inproceedings{zhao2026beyond,
  title = {{Beyond Retrieval: Bi-Temporal State Arbitration for Longitudinal Healthcare Agents}},
  author = {Jianing Zhao, Xiaoquan Zhi and Xinqiang Yu},
  booktitle = {Proceedings of the 4th Workshop on Towards Knowledgeable Foundation Models (KnowFM 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.knowfm-1.10},
  url = {https://doi.org/10.18653/v1/2026.knowfm-1.10}
}

References

15 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Prateek Chhikara, Deshraj Khant, Saket Aryan, and Taranjeet Singh. 2025. Mem0: Building production-ready AI agents with scalable long-term memory. arXiv preprint arXiv:2504.19413.
  2. Cheng Yang, Chuan Zhou, Yixin Xiao, Shen Dong, and Linyuan Zhuang. 2026. Graph-based agent memory: Taxonomy, techniques, and applications. arXiv preprint arXiv:2602.05665. — [arXiv:2602.05665](https://arxiv.org/abs/2602.05665) find
  3. Jon Doyle. 1979. A truth maintenance system. Artificial Intelligence, 12(3):231–272. find
  4. Shanghua Gao, Ada Fang, Yepeng Huang, and 1 others. 2024. Empowering biomedical discovery with AI agents. Cell, 187(22):6125–6151. find
  5. Peter Gärdenfors. 1988. Knowledge in flux: Modeling the dynamics of epistemic states. MIT Press. find
  6. Zhuohan Ge, Haoyang Li, Yubo Wang, Nicole Hu, Chen Jason Zhang, and Qing Li. 2026. ClinicalA-gents: Multi-agent orchestration for clinical decision making with dual-memory. arXiv preprint arXiv:2603.26182. — [arXiv:2603.26182](https://arxiv.org/abs/2603.26182) find
  7. Christian S. Jensen, Michael D. Soo, and Richard T. Snodgrass. 1994. Unifying temporal data models via a conceptual model. Information Systems, 19(7):513–547. find
  8. Jiale Kang, Mingyu Ji, Zhiyuan Zhao, and Tao Bai. 2025. Memory OS of AI agent. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing.
  9. Patrick Lewis, Ethan Perez, Aleksandra Piktus, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. ArXiv:2005.11401.
  10. Ruoqi Li, Xin Wang, Dan Berlowitz, and 1 others. 2025. CARE-AD: A multi-agent large language model framework for Alzheimer’s disease prediction using longitudinal clinical notes. npj Digital Medicine. find
  11. Fang Liu and 1 others. 2025. A foundational architecture for AI agents in healthcare. npj Digital Medicine. PMC12629813. find
  12. Charles Packer, Vivian Fang, Shishir G. Patil, Kevin Lin, Sarah Wooders, and Joseph E. Gonzalez. 2023. MemGPT: Towards LLMs as operating systems. In Advances in Neural Information Processing Systems. ArXiv:2310.08560. find
  13. 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.
  14. Karan Singhal, Shekoofeh Azizi, Tao Tu, and 1 others. 2025. Toward expert-level medical question answering with large language models. Nature Medicine. find
  15. Richard T. Snodgrass and Ilsoo Ahn. 1985. A taxonomy of time databases. ACM SIGMOD Record, 14(4):236–246. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhao-2026-beyond-retrieval-bi-temporal-state/