Submitted toAAMAS 2027Preprint, October 2026

What Is This Memory Useful For?

Measuring Task-Conditioned Utility in Long-Term Agent Memory

Chih-Hsuan Yang1*, Anjir Ahmed Chowdhury2*, Feng Yan2, Rajeev Thakur1

1 Argonne National Laboratory2 University of Houston* Equal contribution

Left: the protocol. A page-visit memory is written once by a task-blind writer and frozen; a frozen reader answers the same later task with identical filler, with and without the memory, and the reward difference is Delta U; repeating this gives a task-wise utility profile with a cost ledger. Right, top: answer-bearing minus matched-control Delta U is positive in the original run (+0.33), a new session with random position (+0.50) and under a second reader (+0.60). Right, bottom: Raw gains about +0.35 on answer-bearing cells, while Summary and Structured stay near zero, at their own length and at a fixed total context. View full-size figure ↗
A memory has no single usefulness. Its effect depends on the task it meets.Left: the paired with/without-memory protocol. Right: answer-bearing minus matched-control ΔU across runs, and bearing-cell ΔU by representation; 95% intervals.

What is a stored memory useful for?

A writer stores each past page visit without ever seeing a future task; later, a frozen reader answers each question twice with identical context, once with the memory and once without it. On LongMemEval-V2 factual-recall tasks, the effect is strongly task-conditioned: within the same memory, answer-bearing task pairs gain +0.33 more reward than matched controls, and the contrast stays positive across sessions, positions and readers. Task-blind summaries and structured records gain less than the raw text, even at a fixed total context, and generic retrieval finds answer-bearing memories far above chance but misses most of them.

Chih-Hsuan Yang, Anjir Ahmed Chowdhury, et al. Preprint, 2026.
paper.pdf

BibTeX
@misc{yang2026memoryuseful,
    title = "What Is This Memory Useful For? Measuring Task-Conditioned Utility in Long-Term Agent Memory",
    author = "Yang, Chih-Hsuan  and
      Chowdhury, Anjir Ahmed  and
      Yan, Feng  and
      Thakur, Rajeev",
    month = oct,
    year = "2026",
    note = "Preprint"
}