Does tool-mediated memory actually hold up?

This is a measurable question, so CueCrux measures it in the open. ScoreCrux is the public benchmark for agent memory, and its findings are the evidence behind the design.

At scale, stuffed context fails. At two million tokens of history, recall on stuffed-context approaches collapsed from 44% to 28% on one model and from 28% to 8% on another. Tool-mediated memory over the same material held 80-100%. The failure mode is structural: past a certain scale, a context window is where information goes to get lost.

Quality does not require volume. In head-to-head answering on the open ScoreCrux benchmark, tool-mediated memory matched vendor-native answer quality on a fraction of the context tokens, which at fleet scale is the difference between a viable bill and an absurd one.

Budgets do real work. Every CueCrux retrieval call carries a hard token_budget and the daemon trims to fit. Under ScoreCrux, budgeted retrieval returned far fewer tokens than naive top-K without giving up recall.

The benchmark is open so the claims stay honest: same corpus, same questions, reproducible runs. If a future release regresses, ScoreCrux is where it will show.