Preprint measures cost of memory use in multi-agent AI workflows

A preprint reports that text retrieved from memory accounts for about 12 percent of the full billed cost in a 200-task benchmark of multi-agent AI workflows. The researchers counted those input tokens directly across real model APIs, then reduced injected tokens by 28.7 percent by shrinking the retrieval window from 32 entries to 2, with accuracy changing only within variation across test runs. The result gives teams a measurable way to manage one part of their AI bills. The study held model tier fixed, found its graph changes roughly cost-neutral on their own, and did not evaluate prompt caching.