AuthorsAsim Biswal, Chuan Lei, Xiao Qin et al.
TL;DR
AgentSM uses structured semantic memory plus composite tools to reuse Text-to-SQL trajectories, reaching 44.8% execution accuracy on Spider 2.0 Lite (+16.1 points over SpiderAgent).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing Text-to-SQL agents repeatedly explore the same schemas, leading to long trajectories and only 25% execution accuracy with 22.62 steps when reading raw trajectories.
On enterprise-style benchmarks like Spider 2.0, this instability and redundancy cause high latency, token bloat, and frequent failures to generate valid SQL for complex databases.
HOW IT WORKS
AgentSM centers on a planner agent, schema linking agent, trajectory store, and composite tools connected through structured trajectory synthesis and retrieval.
You can think of AgentSM like a developer using a well-organized notebook and macros instead of retyping the same exploratory commands for every new query.
This design lets AgentSM reuse prior reasoning paths and bundled tool sequences, enabling targeted exploration and SQL generation that a plain context window or scratchpad cannot sustain across tasks.
DIAGRAM
This diagram shows how AgentSM routes a new Text-to-SQL question through trajectory retrieval, schema linking, and composite tools during inference.
DIAGRAM
This diagram shows how AgentSM is evaluated on Spider 2.0 Lite and how ablations disable trajectory reading or composite tools.
PROCESS
Trajectory Synthesis
AgentSM generates synthetic questions per schema and runs agents with SQL and file tools to populate the trajectory store with exploration rich traces.
Step Classification
AgentSM classifies each trajectory step into exploration, query execution, or validation phases, then restructures them into readable markdown segments.
Trajectory Retrieval
At inference, AgentSM filters trajectories by database, selects the most similar question using embeddings, and loads only the relevant trajectory phase.
Composite Tool Construction
AgentSM mines frequent tool subsequences like get_ext then get_ddl and bundles them into composite tools that the planner agent can call in one step.
KEY CONTRIBUTIONS
AgentSM agentic framework
AgentSM introduces a two agent architecture with a planner agent and schema linking agent plus a trajectory store tailored for enterprise Text-to-SQL workloads like Spider 2.0.
Structured semantic memory
AgentSM designs a structured semantic memory that encodes prior trajectories with step classification and markdown headers, enabling retrieval that cuts steps by 25% and boosts accuracy by 35%.
Composite tools for efficiency
AgentSM automatically constructs composite tools from frequent tool sequences, reducing average steps from 20.36 to 15.99 and latency from 523.7s to 252.4s in ablations.
RESULTS
EX (%)
44.8%
+16.1 over SpiderAgent (claude-3-7-sonnet)
Avg. Steps
16.4
-2.5 vs SpiderAgent (claude-3-7-sonnet)
Avg Latency (s)
247.1
-116.1 seconds vs SpiderAgent (claude-3-7-sonnet)
EX (%) gold tables
57.6%
+12.8 points over AgentSM without gold tables
On Spider 2.0 Lite, which stresses multi dialect, long context Text-to-SQL, AgentSM with Claude 4 Sonnet reaches 44.8% execution accuracy. This demonstrates that structured semantic memory and composite tools let AgentSM trade redundant exploration for more effective reasoning under strict step and latency budgets.
BENCHMARK
On Spider 2.0 Lite, which stresses multi dialect, long context Text-to-SQL, AgentSM with Claude 4 Sonnet reaches 44.8% execution accuracy. This demonstrates that structured semantic memory and composite tools let AgentSM trade redundant exploration for more effective reasoning under strict step and latency budgets.
BENCHMARK
Overall execution accuracy (%) on Spider 2.0 Lite across agentic Text-to-SQL systems.
BENCHMARK
Execution accuracy (%) over 75 sampled Spider 2.0 Lite questions under different AgentSM configurations.
KEY INSIGHT
Reading raw, unstructured trajectories in AgentSM yields no gain over not using any trajectory at all, both stuck at 25% accuracy.
This is surprising because more past information usually helps, but AgentSM shows that unstructured logs just add noise without structure and annotations.
WHY IT MATTERS
AgentSM unlocks reusable, interpretable semantic memory for Text-to-SQL agents, turning one off explorations into durable programs for future queries.
Builders can now design agentic systems that get faster and more accurate with use, instead of re paying the same exploration cost on every enterprise database question.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…