Tool graph memory rarely improves multi-step language model tool use
Rethinking the Role of Tool Graph Memory in Multi-Turn Tool Use
Computation and LanguageArtificial Intelligence
Summary
Using big language models to solve complex tasks often involves calling external tools step by step. The authors looked at a memory method that tries to remember past tool use to pick the next tool better. They found that this memory method usually doesn't help much and can even waste a lot of extra work. Instead, turning past experience into simple text instructions worked better and used fewer resources. This means it’s better to guide the model softly rather than telling it exactly what to do next.
What this means in practice
- •For ai application developers: Improve multi-step AI workflows by converting past tool interactions into textual hints rather than strict tool selections.
- •For automation engineers: Reduce unnecessary resource use in automated assistants by avoiding complex tool memory structures and using simpler experience summaries.
Authors
Uiji Hwang, Richeng Xuan, Taeuk Kim
Abstract
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. 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. The memory's guidance is often incorrect, and random or no guidance retains most of the gain, implying that the benefit stems from extra computation rather than the memory itself. Moreover, even with the correct next tool suggested, many failures remain, such as giving up without a tool call. Storing the same experience instead as textual insights outperforms tool graph memory with almost no extra LLM calls. This suggests that memory serves better as a soft prior than as a prescription of the next tool, leading to a more effective and generalizable solution.