Elamur Research
☰
home / papers / 2025 / Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory

Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv — non-exclusive license to distribute" does not allow redistributing the text itself. Original — arxiv.org/pdf/2504.19413

Summary

LLM-based agents lose track of facts once a conversation outgrows the context window. Longer windows only postpone the problem: history keeps growing, and relevant details get buried among unrelated content. The authors propose Mem0, a memory layer that works incrementally on each new message pair. An LLM first pulls candidate facts from the exchange, using a periodically refreshed conversation summary and a window of recent messages as context. Each candidate is then compared with the most similar stored memories, and the LLM chooses through a tool call whether to add, update, delete or leave the store unchanged. A second variant, Mem0g, stores memories as a directed labeled graph of entities and relations, which helps with relational and temporal questions. On the LOCOMO benchmark, both systems beat memory-augmented baselines, RAG configurations, open-source and proprietary memory tools, and OpenAI's memory feature. They also cost far less than feeding the whole conversation to the model. This shows that structured persistent memory can be accurate and cheap enough for production agents.

Key points

  • Mem0's update step relies on the LLM itself, not a separate classifier, to select one of four operations: ADD, UPDATE, DELETE or NOOP.
  • On LOCOMO, Mem0 gains about 26% relative on the LLM-as-a-Judge (J) metric over OpenAI. Mem0g scores roughly 2% higher overall than base Mem0, reaching J = 68.44%, behind only the full-context approach.
  • Compared with full-context processing, Mem0 cuts p95 latency by about 91% and token cost by more than 90%. Its search latency is p50 0.148s / p95 0.200s, and total latency is p50 0.708s / p95 1.440s.
  • Average memory footprint per conversation is about 7k tokens for Mem0 and about 14k for Mem0g. Zep's graph exceeds 600k tokens, and the raw conversation is roughly 26k.
  • Relative gains over the best competitor per question type are 5% (single-hop), 11% (temporal) and 7% (multi-hop). Mem0g graph construction finishes in under a minute, while Zep's memories took hours to become reliably retrievable.

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: Other

Cited by

114 papers of our library cite it (of the 140673 whose bibliographies we read).

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2504.19413 · arXiv 2504.19413

@misc{chhikara2025mem0,
  title = {{Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory}},
  author = {Chhikara, Prateek and Khant, Dev and Aryan, Saket and Singh, Taranjeet and Yadav, Deshraj},
  year = {2025},
  doi = {10.48550/arxiv.2504.19413},
  eprint = {2504.19413},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2504.19413}
}

References

29 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Anthropic. Model card and evaluations for claude models. Technical report, Anthropic, February 2025. URL https://www.anthropic.com/news/claude-3-7-sonnet . ↗
  2. Jan Assmann. Communicative and cultural memory. In Cultural memories: The geographical point of view, pages 15–27. Springer, 2011. find
  3. Aydar Bulatov, Yury Kuratov, and Mikhail Burtsev. Recurrent memory transformer. Advances in Neural Information Processing Systems, 35:11079–11091, 2022. find
  4. Prateek Chhikara, Jiarui Zhang, Filip Ilievski, Jonathan Francis, and Kaixin Ma. Knowledge-enhanced agents for interactive text games. In Proceedings of the 12th Knowledge Capture Conference 2023, pages 157–165, 2023. find
  5. Fergus IM Craik and Janine M Jennings. Human memory. 1992. find
  6. Gaurav Goswami. Dissecting the metrics: How different evaluation approaches yield diverse results for conversational ai. Authorea Preprints, 2025. find
  7. Tianyu Guo, Druv Pai, Yu Bai, Jiantao Jiao, Michael Jordan, and Song Mei. Active-dormant attention heads: Mechanistically demystifying extreme-token phenomena in llms. In NeurIPS 2024 Workshop on Mathematics of Modern Machine Learning, 2024. find
  8. Kostas Hatalis, Despina Christou, Joshua Myers, Steven Jones, Keith Lambert, Adam Amos-Binks, Zohreh Dannenhauer, and Dustin Dannenhauer. Memory matters: The need to improve long-term memory in llm-agents. In Proceedings of the AAAI Symposium Series, volume 2, pages 277–280, 2023. find
  9. Zihong He, Weizhe Lin, Hao Zheng, Fan Zhang, Matt W Jones, Laurence Aitchison, Xuhai Xu, Miao Liu, Per Ola Kristensson, and Junxiao Shen. Human-inspired perspectives: A survey on ai long-term memory. arXiv preprint arXiv:2411.00489, 2024. — [arXiv:2411.00489](https://arxiv.org/abs/2411.00489) find
  10. Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. find our card
  11. Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. find our card
  12. Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. A human-inspired reading agent with gist memory of very long contexts. In International Conference on Machine Learning, pages 26396–26415. PMLR, 2024. find
  13. Lei Liu, Xiaoyan Yang, Yue Shen, Binbin Hu, Zhiqiang Zhang, Jinjie Gu, and Guannan Zhang. Think-in-memory: Recalling and post-thinking enable llms with long-term memory. arXiv preprint arXiv:2311.08719, 2023. — [arXiv:2311.08719](https://arxiv.org/abs/2311.08719) find
  14. Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. 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 13851–13870, 2024… find
  15. Bodhisattwa Prasad Majumder, Bhavana Dalvi Mishra, Peter Jansen, Oyvind Tafjord, Niket Tandon, Li Zhang, Chris Callison-Burch, and Peter Clark. Clin: A continually learning language agent for rapid task adaptation and generalization. In First Conference on Language Modeling. Elliot Nelson, Georgios… find
  16. Charles Packer, Vivian Fang, Shishir\_G Patil, Kevin Lin, Sarah Wooders, and Joseph\_E Gonzalez. Memgpt: Towards llms as operating systems. 2023. find
  17. Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: A temporal knowledge graph architecture for agent memory. arXiv preprint arXiv:2501.13956, 2025. find
  18. Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D Manning. Raptor: Recursive abstractive processing for tree-organized retrieval. In The Twelfth International Conference on Learning Representations, 2024. find our card
  19. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634– 8652, 2023. find our card
  20. Prabhjot Singh, Prateek Chhikara, and Jasmeet Singh. An ensemble approach for extractive text summarization. In 2020 International Conference on Emerging Trends in Information Technology and Engineering (ic-ETITE), pages 1–7. find
  21. Arpita Soni, Rajeev Arora, Anoop Kumar, and Dheerendra Panwar. Evaluating domain coverage in low-resource generative chatbots: A comparative study of open-domain and closed-domain approaches using bleu scores. In 2024 International Conference on Electrical Electronics and Computing Technologies (ICE… find
  22. Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. find our card
  23. Joan C Timoneda and Sebastián Vallejo Vera. Memory is all you need: Testing how model memory affects llm performance in annotation tasks. arXiv preprint arXiv:2503.04874, 2025. — [arXiv:2503.04874](https://arxiv.org/abs/2503.04874) find
  24. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-mem: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110, 2025. find our card
  25. Yangyang Yu, Haohang Li, Zhi Chen, Yuechen Jiang, Yang Li, Denghui Zhang, Rong Liu, Jordan W Suchow, and Khaldoun Khashanah. Finmem: A performance-enhanced llm trading agent with layered memory and character design. In Proceedings of the AAAI Symposium Series, volume 3, pages 595–597, 2024. find
  26. Jiarui Zhang. Guided profile generation improves personalization with large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 4005–4016, 2024. find
  27. Jiarui Zhang, Filip Ilievski, Kaixin Ma, Jonathan Francis, and Alessandro Oltramari. A study of zero-shot adaptation with commonsense knowledge. In AKBC, 2022. find
  28. Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501, 2024. find
  29. Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19724–19731, 2024. find