General-purpose AI document assistants (e.g., NotebookLM) increasingly play an important role and are widely adopted across diverse domains. However, they consistently struggle on complicated multimodal documents such as financial reports and scientific papers, where hierarchical structures, complex layouts, and interleaved visual elements carry essential semantics. A key reason is that these assistants lack native multimodal data management capabilities: they typically linearize documents into flat text chunks or page images, discarding the logical hierarchy and cross-modal context that are critical for faithful evidence retrieval and reasoning. To address these limitations, we present MoDora, an interactive, tree-structured multimodal document analysis agent harness. Designed to make the document analysis process transparent and controllable, MoDora introduces an end-to-end user experience through three core functionalities: (1) an automated document ingestion engine that seamlessly transforms unstructured PDFs into a layout-aware component tree, preserving both textual hierarchy and visual elements; (2) an interactive structure visualizer that allows users to intuitively inspect the extracted document hierarchy and refine structural relations via drag-and-drop or natural language commands; and (3) a verifiable multimodal QA interface that supports complex, cross-modal queries with precise bounding-box-level grounding back to the original PDF regions, enabling users to effortlessly trace and verify the underlying evidence. Experimental results on the MMDA benchmark show that MoDora achieves an AIC-Acc of 71.1%, outperforming baselines by over 14%.
Yu-Kai Wu, Bang-Rui Xu, Shao-Lin Yu et al.· Proceedings of the VLDB Endo...· 0 citations
Recent advances in large language models (LLMs) have led to the emergence of autonomous agents as a transformative paradigm for building intelligent AI systems, integrating reasoning, planning, tool use, and interaction capabilities to tackle complex, open-ended tasks. Despite their growing sophistication, most agents remain stateless, lacking the ability to retain, organize, and reuse past experiences, which limits their personalization, long-term adaptation, and lifelong learning. To address this limitation, agentic memory has emerged as a critical area of research, equipping agents with mechanisms to store, retrieve, update, and exploit information from prior interactions, reasoning trajectories, tool executions, and environmental feedback. From a data management perspective, agentic memory shares similarities with database systems but also presents opportunities for integration, enhancing query functionality, bridging structured and unstructured data, and optimizing execution efficiency through historical insights. This synergy between agentic memory and databases amplifies the capabilities of autonomous agents and redefines the role of databases in adaptive, intelligent data systems. In this tutorial, we systematically review agentic memory systems from a data management perspective. We introduce the workflow and architecture of agentic memory systems. We summarize their operators, storage, and optimization techniques. We also highlight open research challenges, aiming to inspire further innovation and progress in this exciting field.
Guoliang Li, Jia-Qi Tian, Xuanhe Zhou· Proceedings of the VLDB Endo...· 0 citations
Database kernels continuously incorporate new built-in functions to support diverse applications. Synthesizing these native functions is highly complex, as it requires identifying multiple internal units, placing them in the correct source files, and reusing specific internal references. Although recent advancement in LLM-based coding agent frameworks (harnesses) exhibit superior capability, they are typically general-purpose and struggle with the massive and highly coupled database kernel repositories. To address this, we demonstrate
DBCooker
, a database kernel-specialized coding agent harness.
DBCooker
offers three core functionalities: (1)
Kernel Dependency Inspector
, which enables users to inspect internal units and cross-file references via hybrid declaration collection, distinctive unit identification, and cross-unit reference analysis; (2)
Interactive Synthesis Planner
, which constructs high-level implementation plans through progressive code synthesis and three-stage code validation, allowing developers to verify and refine plans before execution; and (3)
Progressive Debugging Workspace
, which integrates with database-specific pipelines to provide efficient feedback and adaptive repair, powered by techniques such as
operation-as-tool abstraction
and
memory-augmented orchestration.
Conference attendees will experience how
DBCooker
seamlessly transforms high-level SQL function requests into validated, kernel-ready code for mainstream databases (e.g., SQLite, PostgreSQL, DuckDB).
Wei Zhou, Xuanhe Zhou, Qi-Kang He et al.· Proceedings of the VLDB Endo...· 1 citation
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
