Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·1d ago
AI Research
arXiv CS.AI·May 28·8 min read
When Does Memory Help Multi-Trajectory Inference for Tool-Use LLM Agents?
TL;DR
Large language model agents that use external tools often explore multiple execution paths (trajectories) to solve tasks, but it's unclear when storing and reusing past experiences improves performance. Researchers investigated how memory mechanisms affect multi-trajectory inference in tool-use LLM agents.
✦ Why It Matters
Engineers can optimize LLM agent memory design by matching memory strategies to task structure rather than applying memory uniformly.
Key Takeaways
- Test different memory methods with various inference strategies.
- Use MCTS for better performance with reflection memory.
- Apply within-expansion injection to enhance beam search diversity.
- Consider trajectory length reduction when using atomic fact extraction.
How It Works
The framework decomposes memory methods into two axes: the scope of transfer, which distinguishes between knowledge shared within a single reasoning attempt and across multiple attempts, and the type of content transferred, which can include reflections on past attempts, direct observations, or extracted facts. This allows for a systematic evaluation of how different memory techniques interact with various inference strategies.