Memory injection grows linearly with workflow depth and hits 27.6% of cost at depth six.

Total Cost of Agency decomposes multi-agent spend into base prompt, inference, memory injection, miss penalty and context accumulation, measuring injected tokens with a two-pass non-billable count rather than word-count proxies, which is a number production observability tools don't separate. On a 200-task enterprise benchmark against real model APIs, memory injection is about 12% of billed cost and 13.6% of the variable cost a compile-time optimizer can touch, rising from a structural zero at depth one with R²=0.9974. Cutting retrieval window capacity from 32 entries to 2 dropped injected tokens 28.7% with accuracy change inside seed variance. The authors state plainly that prompt caching wasn't evaluated. (arXiv 2609.23790)