Rethinking the Role of Tool Graph Memory in Multi-Turn Tool Use
Large language model (LLM) agents solve complex tasks by calling external tools over multiple turns.
ProofPaper ↗
Key points
- Recent work has proposed tool graph memory, which stores past experience as tool transitions to guide subsequent selection.
- However, it remains unclear whether this component consistently improves agents' tool use or undermines it.
- We evaluate tool graph memory on nine models and two benchmarks, finding that it seldom improves significantly over baselines, and that its gains, where present, cost over twice the LLM calls.
- Storing the same experience instead as textual insights outperforms tool graph memory with almost no extra LLM calls.
Sources (1)
- [1]Rethinking the Role of Tool Graph Memory in Multi-Turn Tool UsearXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 09:24 AM
Large language model (LLM) agents solve complex tasks by calling external tools over multiple turns.
Recent work has proposed tool graph memory, which stores past experience as tool transitions to guide subsequent selection.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026ISW-Agent: Integrating LLM-Driven Dialogue with n8n Workflow for Internal Solitary Wave Remote Sensing Image Processing
- Oct 15, 2026Incentivizing Tool-Use Reinforcement Learning for Post-Arrival Emergency Care Prediction
- Oct 15, 2026PriorityFEC: Deadline-Aware Streaming FEC for KV-Cache Handoff in Edge Multi-Agent Systems
- Oct 15, 2026JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs
- Oct 15, 2026Sci$Π$vot: Benchmarking Principal-Investigator Judgment at Pivotal Moments of Scientific AutoResearch Agents
- Oct 14, 2026AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety