Skip to content

Research1 min read

MERIT Benchmark Evaluates Long-Term Memory in Tool-Using Agents

A new benchmark, MERIT, measures the cost-effectiveness of long-term memory for LLM agents executing tasks. Results show structured fact stores and LLM summarization outperform embedding retrieval in maintaining task success, with full replay proving uneconomical.

By OpenSmartRoute editorial · written through the router by writer-small

From arXiv cs.AI - “When Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents

The research introduces MERIT, a benchmark and harness designed to assess the utility of long-term memory in tool-using LLM agents. It focuses on measuring the marginal cost of memory operations, providing a more realistic evaluation than traditional conversational recall benchmarks. The benchmark incorporates episodic tool-use tasks across three domains, verifying dependencies on earlier-episode facts through automated leak checks, difficulty ladders, and controlled memory corruption. The system utilizes token and dollar metering for every memory operation, tracking costs precisely.

Experiments across 23,440 scored episodes with a total cost of $42.57, involved a two-generation pilot on gpt-4.1-mini and a 3-model x 3-seed grid. Results demonstrated that structured fact stores and LLM summarization consistently achieved success rates of 70-100% when recalling updated facts, while embedding retrieval performed unpredictably, with a maximum seed gap of 0.45.

Specifically, agents acting on correctly retrieved values only succeeded 55% of the time. The implementation of memory significantly impacted task success, with a swap moving performance by up to 60 points. Full replay was never economically viable, with the best condition delivering 2.7-3.9x its marginal utility per dollar.

The benchmark, harness, and all traces are released. Source: https://arxiv.org/abs/2609.05441v1

Published Sep 9, 2026 · 190 words

Keep reading

Related posts

More in Research

Research1 min read

Reasoning-Aware Compression for LLM Energy Efficiency

New compression methods risk damaging critical reasoning circuits in Large Reasoning Models (LRMs) by applying uniform quantization. This research presents a framework that identifies and protects vulnerable reasoning circuits, achieving significant energy reductions without compromising performance.

Research1 min read

EdgeMem: LLM-Free Agent Memory Construction

EdgeMem is a new agent memory method that preserves original interaction turns using a multi-anchor hypergraph, eliminating the need for LLM-generated representations. Experiments demonstrate strong retrieval and question answering performance without relying on generative LLMs for memory management.

Research1 min read

ARC-Bench: Evaluating Action Ranking in Frozen JEPA World Models

ARC-Bench identifies a fundamental flaw in frozen JEPA-style world models: predicted action rankings based on latent space distances consistently fail to align with true cost. This research introduces a protocol to measure this failure and explains why it remains hidden due to replanning masking.