When Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents

A new benchmark called MERIT (Memory Evaluation for Realistic Instrumented Tasks) shows that long-term memory lifts dependent-task success for tool-using LLM agents from a leak-verified floor of 0.00 to 0.55-1.00, according to an arXiv paper (2609.05441v1). Across 23,440 scored episodes costing $42.57, embedding retrieval on updated facts collapsed unpredictably (0.30-0.95 across models, max seed gap 0.45) and agents acted on a correctly retrieved value only 55% of the time, while update-on-write stores including LLM summarization held at 0.70-1.00. The authors report that swapping a memory's implementation moves task success by up to 60 points and that full replay is never economical, with the best condition per domain delivering 2.7-3.9x its marginal utility per dollar.

arXiv:2609.05441v1 Announce Type: new Abstract: Long-term memory for LLM agents is evaluated today by conversational recall benchmarks LoCoMo, LongMemEval , which measure question answering over dialogue history, not whether remembered facts change what a tool-using agent does. We present MERIT Memory Evaluation for Realistic Instrumented Tasks , a benchmark and harness that measures the marginal utility of memory for task-executing agents under explicit cost accounting. MERIT provides episodic tool-use tasks in three domains whose dependence on earlier-episode facts is verified by an automated leak check; a difficulty ladder ending in updated-fact recall; controlled memory corruption; and full token and dollar metering of every memory operation. Across 23,440 scored episodes $42.57 , a two-generation pilot on gpt-4.1-mini and a preregistered 3-model x 3-seed grid GPT-4.1, Claude Haiku 4.5; memory side held fixed , memory lifts dependent-task success from a leak-verified floor of 0.00 to 0.55-1.00. On updated facts, embedding retrieval collapses unpredictably 0.30-0.95 across models; max seed gap 0.45 , and agents act on a correctly retrieved value only 55% of the time, while update-on-write stores a structured fact store and, notably, LLM summarization remain at 0.70-1.00; the hybrid is worse than the fact store alone. A latest-generation spot-check Claude Sonnet 5, gated on a clean full-replay control reproduces the pattern. Swapping a memory's implementation moves task success by up to 60 points, and full replay is never economical: the best condition per domain delivers 2.7-3.9x its marginal utility per dollar. We release the benchmark, harness, and all traces.