# The SecondOS Benchmark

Our own open, reproducible benchmark for structure-aware codebase memory, measured on five real repos.

- **~5–9× fewer tokens per question** — map cards vs re-reading the same top-K files (secondos 4.7×, gurulu 5.6×, tcg-terminal 8.8×).
- **Bounded orientation map** — ~2–2.6k tokens regardless of repo size (51 → 2,302 files), vs 166–1,689× more for raw source.
- **Code search:** precision@1 100%, MRR 1.000 (34-symbol discriminating fixture, rerank on).
- **Memory recall:** precision@1 92%, MRR 0.960 (48 distractors), plus structural invariants (fresh > stale, expiry/conflict labeled).
- **Graph resolution** on 7 real OSS repos (Python 60 · PHP 57 · Rust 51 · Java 45 · TS 40 · Go 24 · Ruby 9 % of decidable calls) — a precision stance, never guessing.

Why our own benchmark and not LOCOMO/DMR? Those measure conversational memory — a different category. Reproduce ours: `bun scripts/benchmark-context.ts <repo>`.