MARMo: An Evidence-Guided and Memory-Augmented Multi-Agent Repair Framework for Accurate Text-to-NoSQL Parsing

ACL ARR 2026 January Submission7395 Authors

Keywords: applications, code models, LLM/AI agents

Abstract:

Text-to-NoSQL aims to extend natural language interfaces (NLIs) beyond traditional Text-to-SQL, enabling intuitive querying of NoSQL systems such as MongoDB—widely adopted in modern data-intensive applications.
While recent advances in Text-to-NoSQL have yielded benchmark datasets, existing systems still suffer from erroneous query generation and a lack of robust mechanisms for error detection, explanation and iterative correction.
In this paper, we propose MARMo: a novel evidence-guided, memory-augmented multi-agent repair framework for accurate Text-to-NoSQL parsing (focusing on MongoDB).
The framework employs a modular, plug-and-play architecture that coordinates multiple specialized agents to perform iterative error diagnosis and adaptive refinement.
Extensive experiments on the TEND benchmark validate that MARMo, as a modular and interpretable framework, achieves overall improvement on repair accuracy and generalization for Text-to-NoSQL.
Our work not only enhances the reliability of natural language-driven NoSQL query systems but also paves the way for future research into robust, human-in-the-loop NLIs for unstructured and semi-structured data management.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: LLM agents, multi-agent systems, agent memory

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 7395