Research Project on Agentic Memory Management

Picture of mengxuan-zhang.md Mengxuan Zhang

19 Oct 2026

How do AI agents form useful memories, maintain them over time, and use them to guide future decisions? How can we evaluate whether these memories help agents perform better?

We invite students to develop a benchmark for evaluating agent memory, examining what an agent remembers and how effectively it uses that information to complete tasks.

Project Overview

Agent memory enables agents to store and use information for different tasks. Its lifecycle involves three closely connected processes: 1) formation, which transforms observations, intermediate results, and feedback into useful memories; 2) evolution, which maintains those memories through consolidation, updating, and forgetting; and 3) retrieval, which makes relevant information available for reasoning and action.

This two-semester project will begin with a structured literature review of agent memory methods, then implement task environments and evaluation pipelines, reproduce representative baselines, and compare methods under controlled conditions with clearly documented resource budgets. The intended outcome is a reproducible benchmark, accompanied by baseline implementations and an empirical analysis of existing methods’ strengths and limitations. Students will follow relevant research and open-source developments throughout the project to keep its design informed by emerging findings.

What You Can Gain

  • Research mentoring and practical guidance on identifying research problems, designing experiments, analysing results, and writing papers.
  • Opportunities to contribute to high-quality research publications and build experience for further study, including a PhD.
  • Practical skills in agent architectures, memory systems, and information retrieval, which may enhance your future career prospects.

Requirments

  1. Solid AI/ML knowledge, strong Python programming skills, and practical software engineering experience.
  2. Familiarity with agent frameworks and workflows, and the ability to independently develop task environments and evaluation pipelines.
  3. Ability to critically engage with recent research and reproduce, analyse, and document experimental results.
  4. Access, through your institution or research environment, to computing resources and model inference facilities sufficient for reproducing baselines and conducting repeated evaluations.

Hu, Yuyang, et al. “Memory in the age of ai agents.” arXiv preprint arXiv:2512.13564 (2025).

Wu, Di, et al. “Longmemeval: Benchmarking chat assistants on long-term interactive memory.” arXiv preprint arXiv:2410.10813 (2024).

He, Zexue, et al. “Memoryarena: Benchmarking agent memory in interdependent multi-session agentic tasks.” arXiv preprint arXiv:2602.16313 (2026).

If you are interested in this project and meet the listed requirements, please first read the relevant papers, contact Mengxuan, and attach your transcripts via email.

arrow-left bars magnifying-glass xmark