AgentSM: Semantic Memory for Agentic Text-to-SQL

TL;DR

AgentSM introduces a semantic-memory-based, two-agent framework for Text-to-SQL that reuses structured execution trajectories to reduce redundant exploration and improve reliability on large, multi-dialect schemas. By synthesizing and storing interpretable trajectories and coupling frequently co-occurring tool sequences into composite tools, AgentSM achieves substantial efficiency and accuracy gains, including a $44.8\%$ execution accuracy on Spider 2.0 Lite and a $25\%$ reduction in average trajectory length along with a $35\%$ accuracy improvement in ablations. The approach demonstrates scalable, enterprise-ready performance with reduced latency and token usage, while maintaining flexibility across diverse databases and dialects. These results highlight the value of structured memory and modular tool design for robust, agentic reasoning in complex data tasks and suggest broad applicability to other data-centric AI workflows beyond Text-to-SQL.

Abstract

Recent advances in LLM-based Text-to-SQL have achieved remarkable gains on public benchmarks such as BIRD and Spider. Yet, these systems struggle to scale in realistic enterprise settings with large, complex schemas, diverse SQL dialects, and expensive multi-step reasoning. Emerging agentic approaches show potential for adaptive reasoning but often suffer from inefficiency and instability-repeating interactions with databases, producing inconsistent outputs, and occasionally failing to generate valid answers. To address these challenges, we introduce Agent Semantic Memory (AgentSM), an agentic framework for Text-to-SQL that builds and leverages interpretable semantic memory. Instead of relying on raw scratchpads or vector retrieval, AgentSM captures prior execution traces-or synthesizes curated ones-as structured programs that directly guide future reasoning. This design enables systematic reuse of reasoning paths, which allows agents to scale to larger schemas, more complex questions, and longer trajectories efficiently and reliably. Compared to state-of-the-art systems, AgentSM achieves higher efficiency by reducing average token usage and trajectory length by 25% and 35%, respectively, on the Spider 2.0 benchmark. It also improves execution accuracy, reaching a state-of-the-art accuracy of 44.8% on the Spider 2.0 Lite benchmark.

Figures (5)

• Figure 1: The standard workflow for Text-to-SQL agents is a series of steps alternating between data exploration, query generation, and answer validation.
• Figure 2: Distribution of the first seven steps in agent trajectories on the firebase from Spider 2 lei2024spider, reporting the percentage of instances in which each specific action is taken.
• Figure 3: An overview of the AgentSM architecture that leverages trajectories in structured semantic memory.
• Figure 4: Analysis of tool usage and composite tool construction. (a) shows that most steps use zero or one tool, motivating composition. (b) illustrates how frequently co-occurring tools are merged into a higher-level composite tool.
• Figure 5: Trajectory composition: median steps split into exploration vs. execution and validation.