T5. Cost and the search tax

Without memory, an assistant that needs a specific project fact has to go find it: it greps and reads files across many turns before it can answer. With the fact already in context, it answers on the first turn. We measured that gap with real headless sessions run against an actual repository, memory off versus memory on, over 8 tasks and 5 runs each.

The unaided session pays a search tax to reach the same correct answer:

• ~5x more turns and ~4.6x more input tokens
• ~5x slower to a usable answer
• cheaper per query with memory, by a factor that ranges from about 2x (a cold session) to roughly 8x (a warm one, where prompt caching is in effect)

The size of the tax scales with how buried the fact is: a deeply buried value costs the unaided agent well over ten times the turns, while an easy-to-find constant costs only a little more. We report the per-task spread rather than a single averaged number.