DolphinBench tests agent memory on 600 tasks

DolphinBench evaluates agent memory through 600 verified knowledge-work tasks: 200 each across three personas, with roughly 500,000 tokens of history per persona. Its arXiv paper requires evaluations to report accuracy alongside total cost and latency, rather than rewarding recall at any price. The release gives builders a cleaner way to compare memory layers against their operating budget; pair it with the archive’s analysis of shared agent memory before buying a larger context window as a substitute for durable memory.