What Is This Memory Useful For? Measuring Task-Conditioned Utility in Long-Term Agent Memory

| Source: arXiv AI

Tags: Agent Memory, LongMemEval-V2, Memory Compression, LLM Agents, Retrieval

A study of long-term agent memory measures task-conditioned utility and finds memories help most when they hold answer-bearing information, with raw storage beating task-blind summaries.

Details

Agent memory systems often assign each stored item one importance score, but a memory that answers one future question may not help another. The authors measure task-conditioned utility: a task-blind writer stores a memory, then a frozen reader answers a later task with and without it, using identical filler, while a cost ledger tracks resources used. On factual-recall tasks built from LongMemEval-V2 web-agent memories, the mean paired reward gain on answer-bearing pairs exceeded matched controls by +0.33 (95% interval +0.21 to +0.46). Task-blind summaries and structured records gave smaller gains than raw memories, even at fixed context size. The effect held in a new session with randomized memory position and under a second reader, a caution for teams that compress agent memory aggressively.