Sign InOpen Brain
arXivPaperNeeds Review

DolphinBench: Mapping the Pareto Frontier of Agent Memory

DolphinBench tests whether agent memory improves real task completion, while requiring accuracy, cost and latency so retrieval quality cannot hide impractical tradeoffs.

arXiv · Sep 21, 2026
Open Source Open MarkdownOpen JSON
Source Summary

DolphinBench uses **three knowledge-work personas**, each with roughly **500k tokens** of message history and **200 tasks**. Every task is checked by requiring an agent to pass with the relevant history and fail without it.

Practical Implication

Use this shape of evaluation for agent memory: test downstream work rather than prompted recall, verify that history is causally useful, and report **accuracy, total cost and latency** together. The dataset and evaluation code are available from the project.

Agent-Ready Context
DolphinBench uses **three knowledge-work personas**, each with roughly **500k tokens** of message history and **200 tasks**. Every task is checked by requiring an agent to pass with the relevant history and fail without it.

Use this shape of evaluation for agent memory: test downstream work rather than prompted recall, verify that history is causally useful, and report **accuracy, total cost and latency** together. The dataset and evaluation code are available from the project.

The material does not report baseline scores or show how representative the personas are of software work. Its history/no-history filter strengthens task validity, but it does not by itself isolate which memory component produced a gain.
Connected Context · Feed7 Judgment

This makes memory evaluation more defensible by requiring each task to depend causally on supplied history and by treating accuracy, cost and latency as a joint frontier. It narrows generic recall tests to downstream knowledge work, but does not identify which memory subsystem caused an improvement or show that three synthetic personas represent software-agent workloads.

Vending-Bench: Long-Horizon Agent Evals — Lukas Petersson, Andon LabsBoth emphasize long-horizon evaluation, but DolphinBench adds a causal history/no-history check while Vending-Bench adds the warning that simulated and real-world behavior can diverge.QuoteBench: How Matched Scores Can Hide Command-Path FailuresQuoteBench shows why DolphinBench’s reported frontier must be attributed to the full memory-and-harness execution path, especially because DolphinBench does not isolate individual memory components.PAIChecker: Uncovering and Checking PR-Issue Misalignment in SWE-Bench-Like BenchmarksPAIChecker’s prompt–oracle alignment test parallels DolphinBench’s history-dependence filter: both validate that an instance actually measures the capability it claims before scores are trusted.Change2Task: From Repository Changes to Executable Coding Agent Tasks and EnvironmentsChange2Task could supply renewable software-work tasks for extending DolphinBench beyond its three personas, while its own alignment and workload-coverage requirements would still need separate validation.
Context Map
benchmark#agent-memory#agent-evals#benchmark-integrity
Uncertainty
The material does not report baseline scores or show how representative the personas are of software work. Its history/no-history filter strengthens task validity, but it does not by itself isolate which memory component produced a gain.