Elamur Research⌕
☰
home / papers / 2026 / Bounded Conversational Memory with Hybrid Retrieval and Evidence Highlighting for Multi-Session Dialogue Systems

Bounded Conversational Memory with Hybrid Retrieval and Evidence Highlighting for Multi-Session Dialogue Systems

Summary

LLM-based dialogue systems do not remember earlier sessions without an external memory, and keeping every past turn eventually overloads the context window. Most memory systems treat this as a retrieval problem. The authors test whether the way retrieved material is structured and shown to the model also affects accuracy. They propose the Context Persistence Protocol (CPP), a modular memory layer with four steps:

  • splitting dialogue into topical segments of 3–10 turns,
  • indexing each segment twice (dense BGE-M3 vectors plus BM25),
  • retrieving under a fixed storage cap, using RRF fusion and cross-encoder reranking,
  • marking the single most query-relevant line of each segment as "Evidence".

On LoCoMo and LongMemEval, CPP's session-level retrieval scores equal or slightly trail those of an unbounded flat-turn baseline, yet CPP answers many more questions correctly. The gains are largest on multi-hop questions. This suggests that context presentation, especially evidence highlighting, drives accuracy somewhat independently of retrieval coverage. That matters for designers of long-term assistants, who may get more from organizing context than from storing more of it.

Key points

  • LoCoMo results: CPP reaches 81.4% multiple-choice accuracy. That is 3.4 pp above the unbounded flat baseline overall and 15.9 pp above it on multi-hop questions, while indexing 4.8× fewer retrieval units.
  • LongMemEval results: at k=15, CPP scores 0.606 against the baseline's 0.360. The baseline degrades when given 15 noisy raw turns, while CPP stays fairly stable. In a development-set ablation, evidence highlighting was the largest single contributor, at +26.3 pp.
  • Components: BGE-M3 embeddings (dimension 1024), BM25, RRF (k=60) and the ms-marco-MiniLM-L-6-v2 cross-encoder. A rule-based pass also rewrites first-person pronouns into speaker names.
  • Storage cap: the store is limited to B=400 segments with FIFO eviction. The benchmarks used only 27–31% of this cap, so eviction never triggered and was not actually tested.
  • Limitations:
    • Only English data was used.
    • The Qwen2.5-14B judge comes from the same model family as the generator and was not checked against human annotations.
    • CPP passes about 3.4× more context words to the model than the baseline, and no token-matched comparison has been run yet.

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

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.sigdial-1.55

@inproceedings{mudunuri2026bounded,
  title = {{Bounded Conversational Memory with Hybrid Retrieval and Evidence Highlighting for Multi-Session Dialogue Systems}},
  author = {Mudunuri, Manasa and Guntupalli, Chandra Kiran and Pokharel, Shiraj and Patterson, Murray and Kubek, Mario M.},
  booktitle = {Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue},
  year = {2026},
  doi = {10.18653/v1/2026.sigdial-1.55},
  url = {https://doi.org/10.18653/v1/2026.sigdial-1.55}
}

References

21 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024a. M3-Embedding: Multilinguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. In Findings of the Association for Computational Linguistics: ACL 2024, pages 2318–2335, Bangk… find
  2. Tong Chen, Hongwei Wang, Sihao Chen, Wenhao Yu, Kaixin Ma, Xinran Zhao, Hongming Zhang, and Dong Yu. 2024b. Dense X retrieval: What retrieval granularity should we use? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 15159–15177, Miami, Florid… find
  3. Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. Mem0: Building production-ready AI agents with scalable long-term memory. In ECAI 2025 — 28th European Conference on Artificial Intelligence, volume 413 of Frontiers in Artificial Intelligence and Applications, pages… find our card
  4. Gordon V. Cormack, Charles L. A. Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 758–759, Boston, MA, USA.… find our card
  5. Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Naik, Pengshan Cai, and Alfio Gliozzo. 2022. Re2G: Retrieve, rerank, generate. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages… find
  6. CPP is designed to assist users of long-term conversational AI systems, but its deployment raises privacy considerations that should be acknowledged. Storing multi-session conversation history, even in bounded compressed form, involves retaining personal information that users may have shared 793 find
  7. Maarten Grootendorst. 2022. BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv preprint, arXiv:2203.05794. on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992, Hong Kong, China. Association for Computational Linguistics. find our card
  8. Jiazheng Kang, Mingming Ji, Zhe Zhao, and Ting Bai. 2025. Memory OS of AI agent. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 25961–25970, Suzhou, China. Association for Computational Linguistics. find
  9. Stephen Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval, 3(4):333–389. find our card
  10. Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769–6781, Onli… find our card
  11. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Pr… find
  12. 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 The Thirteenth International Conference on Learning Representations. find
  13. Zhaofen Wu, Hanrong Zhang, Fulin Lin, Wujiang Xu, Xinran Xu, Yankai Chen, Henry Peng Zou, Shaowen Chen, Weizhi Zhang, Xue Liu, Philip S. Yu, and Hongwei Wang. 2026. GAM: Hierarchical graph-based agentic memory for LLM agents. arXiv preprint, arXiv:2604.12285. find
  14. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2025. A-MEM: Agentic memory for LLM agents. In Advances in Neural Information Processing Systems, volume 38. Curran Associates, Inc. find our card
  15. Jiaqi Liu, Yaofeng Su, Peng Xia, Siwei Han, Zeyu Zheng, Cihang Xie, Mingyu Ding, and Huaxiu Yao. 2026. SimpleMem: Efficient lifelong memory for LLM agents. arXiv preprint, arXiv:2601.02553. — [arXiv:2601.02553](https://arxiv.org/abs/2601.02553) find
  16. Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173. find our card
  17. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, and 11 others. 2024. Qwen2.5 technical report. arXiv preprint, arXiv:2… find
  18. Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. 2024. Evaluating very long-term conversational memory of LLM agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13851– 1387… find
  19. 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. find
  20. Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Xufang Luo, Hao Cheng, Dongsheng Li, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Jianfeng Gao. 2025. SeCom: On memory construction and retrieval for personalized conversational agents. In The Thirteenth International Conference on Learning Represe… find
  21. Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/mudunuri-2026-bounded-conversational-memory-hybrid-retrieval/