> For the complete documentation index, see [llms.txt](https://sjtu-edu-4.gitbook.io/sjtu.edu-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://sjtu-edu-4.gitbook.io/sjtu.edu-docs/readme.md). # Hongming Xu (许鸿铭) [**GitHub**](https://github.com/Homy-Xu) · [**CV**](https://github.com/Homy-Xu/homepage/blob/main/files/CV_Hongming_Xu.pdf) · **Email:** · **Undergraduate Student, School of Computer Science, Shanghai Jiao Tong University**\ **Information Security · Zhiyuan Honors Program** *** I am an undergraduate student at the School of Computer Science, Shanghai Jiao Tong University, and a member of the Zhiyuan Honors Program. My research journey began with **database systems and data management**, where I studied how large language models interact with structured data and heterogeneous database systems. This experience led me to view memory as a data-management problem and to explore how memory for intelligent agents should be represented, stored, retrieved, updated, and maintained. In *Are We Ready For An Agent-Native Memory System?*, I studied agent memory from this perspective through system decomposition and systematic evaluation. Building on this foundation, my current research focuses on **coding foundation models, long-horizon agents, and memory for self-improving AI systems**. At MemTensor, I contribute to coding-oriented post-training of a large-scale foundation model and study memory mechanisms for long-horizon coding agents. I am especially interested in **agent memory, parametric memory, coding-model post-training, and recursive self-improvement**—in particular, how models and agents can accumulate, update, consolidate, and reuse experience over long-term interaction. #### Research Interests * **Coding Foundation Models & Agentic Coding** — coding-model post-training, software-engineering task synthesis, trajectory curation, repository-level reasoning, and autonomous software engineering * **Agent Memory & Long-Horizon Agents** — persistent execution memory, state tracking, retrieval, validation, consolidation, and memory maintenance over long executions * **Parametric Memory & Model Adaptation** — storing and updating experience in model parameters, continual adaptation, and the interaction between parametric and external memory * **Self-Improving & Recursive Self-Improving AI** — learning from execution traces and failures, capability-gap discovery, iterative data/model improvement, and long-term self-improvement * **Data Systems for Intelligent Agents** — data-management abstractions for agent memory, structured state, retrieval, update, and maintenance *** #### Selected Research **MemTrace: State-Consistent Memory for Long-Horizon Coding Agents** **First Author · Aug 2026 – Present · Under Review at ICLR 2027**\ **Coding Agents / Long-Horizon Agents / Agent Memory** [Paper](https://arxiv.org/pdf/2610.04838)· [Repository](https://github.com/Homy-Xu/MemTrace) Long-horizon coding agents accumulate large amounts of execution history while the underlying repository continues to evolve. **MemTrace** formulates this challenge as a **state–memory alignment problem**: historical evidence should not only be recoverable, but should remain valid with respect to the current task and repository state. MemTrace stores execution evidence as immutable **Memory Traces** with explicit repository provenance. It combines a **Memory Trace Graph (MTG)**, which organizes dependencies, verification, correction, and supersession relations across execution history, with a **Repository State Graph (RSG)**, which represents the current repository structure. Trace-to-entity bindings connect the two views, allowing historical evidence to be validated before reuse. **Highlights** * Evaluated on **DeepSWE, SWE-EVO, and SWE-Milestone** under two coding-agent harnesses with the same backbone model. * MemTrace consistently improves the primary metric over the corresponding Native memory strategy across the completed benchmark × harness settings. * Trajectory analysis shows that **57.0% of file-bound Memory Traces** become exposed to later repository modifications, highlighting the prevalence of state–memory misalignment in long-horizon coding. *** **Are We Ready For An Agent-Native Memory System?** **Under Review at VLDB · Agent Memory / Data Management Systems**\ [Paper](https://arxiv.org/abs/2606.24775) · [Code](https://github.com/OpenDataBox/MemoryData) **Wei Zhou, Xuanhe Zhou, Shaokun Han, Hongming Xu, Guoliang Li, Zhiyu Li, Feiyu Xiong, Fan Wu** This work studies **LLM agent memory from a data-management perspective**, treating memory as a system composed of multiple interacting modules rather than a monolithic retrieval component. It decomposes agent memory into core components including **representation and storage, extraction, retrieval and routing, and maintenance**, and systematically evaluates heterogeneous memory architectures across effectiveness, retrieval fidelity, robustness to dynamic updates, long-horizon stability, and runtime cost. **My contributions** * Participated in decomposing agent memory into core modules including **representation and storage, extraction, retrieval/routing, and maintenance**. * Contributed to systematic evaluation across task effectiveness, retrieval fidelity, robustness to dynamic updates, long-horizon stability, and runtime cost. * Helped analyze experimental results and identify where different memory architectures are effective across conversational QA, factual recall, temporal reasoning, and long-running scenarios. *** **Dial: A Knowledge-Grounded Dialect-Specific NL2SQL System** **PVLDB 2026 · Co-first Author (2nd)**\ [Paper](https://arxiv.org/abs/2603.07449) · [Code](https://github.com/OpenDataBox/Dial) **Xiang Zhang, Hongming Xu, Le Zhou, Wei Zhou, Xuanhe Zhou, Guoliang Li, Yuyu Luo, Changdong Liu, Guorun Chen, Jiang Liao, Fan Wu** Dial studies **dialect-specific NL2SQL** for heterogeneous database systems, where different SQL dialects introduce distinct syntax, functions, and execution constraints. The work explores how natural-language requests can be transformed into dialect-aware intermediate representations and combined with knowledge retrieval to improve SQL generation across heterogeneous database systems. **My contributions** * Contributed to natural-language request preprocessing and the construction of **NL-LQP / dialect-aware logical plans**, supporting semantic decomposition from user requests to logical plans. * Contributed to dialect-sensitive intermediate representations, operator annotation, and knowledge retrieval for error-prone SQL semantics such as dates, strings, and type conversions. * Participated in the design and initialization of the **HINT-KB** knowledge base, as well as multi-round feedback generation, semantic correction, failure analysis, and rebuttal revisions. *** **Tensor-Network Physical Pages for Quantum-Ready Data Management** **Accepted at QCDKM 2026 · Quantum Database / Approximate Data Management** This work explores **tensor-network physical-page representations** for quantum-ready data management and approximate data processing. **My contributions** * Participated in research on **TN-Page**, a tensor-network-based physical page representation for approximate data management and data-loading abstractions targeting quantum backends. * Contributed to workload-aware bucketization, semantic tensorization, tensor-network factorization, and backend-aware encoding. * Helped organize evaluation around fidelity, compression, qubit usage, and gate depth. *** #### Research Experience **MemTensor** **Research Intern · Mentor: Juncheng Zhang (Tsinghua University)**\ **Aug 2026 – Present** * Contribute to coding-oriented post-training of a **284B-parameter DeepSeek-V4-Flash-0731** foundation model. Design and synthesize software-engineering tasks, curate and filter training data and execution trajectories, and evaluate post-trained checkpoints on coding and repository-level tasks. * Analyze capability gains, failure modes, and remaining skill gaps across training iterations, using evaluation results to guide subsequent **data selection, task construction, trajectory filtering, and post-training iterations**. * Led the first-author **MemTrace** project, studying state-consistent memory for repository-scale coding tasks and the interaction between execution memory and evolving repository state. * Explore memory mechanisms for long-horizon coding agents, with a focus on **persistent execution history, repository-state alignment, and reliable reuse of historical evidence**. **Shanghai Jiao Tong University — Theseus Lab** **Undergraduate Researcher · Supervisor: Prof. Xuanhe Zhou**\ **Aug 2025 – Jun 2026** * Worked on **Dial**, a knowledge-grounded dialect-specific NL2SQL system, and contributed to dialect-aware logical planning and the construction of **HINT-KB**. * Participated in knowledge organization and initialization, retrieval components, semantic correction, iterative feedback generation, and failure analysis for heterogeneous SQL dialects. * Extended this data-management perspective toward **agent memory and memory management** through *Are We Ready For An Agent-Native Memory System?*, studying memory representation and storage, extraction, retrieval/routing, maintenance, and systematic evaluation. * This line of work motivated my transition from database systems toward **agent memory, long-horizon systems, and eventually coding-agent memory**. *** #### News * **Sep 2026** · **MemTrace**, our work on state-consistent memory for long-horizon coding agents, is **under review at ICLR 2027**. * **Aug 2026** · Started my research internship at **MemTensor**, working on coding-model post-training and long-horizon agent memory. * **Jul 2026** · Started researching on **MemTrace** and state-consistent memory for long-horizon coding agents. * **July 2026** · *Dial: A Knowledge-Grounded Dialect-Specific NL2SQL System* is accepted by **PVLDB 2026**. * **Jun 2026** · *Are We Ready For An Agent-Native Memory System?* is released on arXiv. * **April 2026** · Started research on agent memory at **Theseus Lab**. * **March 2026** · *Dial: A Knowledge-Grounded Dialect-Specific NL2SQL System is* released on arXiv. * **Sep 2025** · Started research on database systems and NL2SQL at **Theseus Lab**. *** #### Education **Shanghai Jiao Tong University** **School of Computer Science · Information Security**\ **Zhiyuan Honors Program**\ *Sep 2024 – Present · Shanghai, China* *** #### Honors * **Zhiyuan Honors Scholarship** --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://sjtu-edu-4.gitbook.io/sjtu.edu-docs/readme.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.