Unified Framework for Memory in Multi-Trajectory Tool-Use LLM Agents

A recent preprint on arXiv (2605.28224) introduces a comprehensive framework for managing memory in multi-trajectory inference for tool-utilizing LLM agents. This framework categorizes memory into two dimensions: the scope of transfer (whether it is within an expansion or across trajectories) and the level of abstraction of the transferred information. It assesses four memory techniques using three inference approaches (best-of-N, beam search, MCTS) across four benchmarks related to tool use, including SQL, knowledge-graphs, and CLI environments, all in a verifier-free context. The research seeks to determine if the observed improvements are due to the characteristics of memory abstraction or the inference method, as current techniques have only been tested with single inference strategies on individual tasks.