External benchmark scores are pending, and we say so
Published LoCoMo and LongMemEval scores require an LLM-judge protocol; our judged runs are pending and we won't publish heuristic floors as results. Field-leading reference scores (re-verified 2026-08-16: mem0 LoCoMo 92.5, Zep DMR 98.2) come from hosted stacks with answer models in the loop — a different protocol than air-gapped local retrieval, so we never mix columns.
Every claim states its protocol
Hardware, dataset, isolation proof, and artifact filenames are recorded per run. Consolidation A/B: on a duplicate-noisy stream, recall@8 +1.25 pt, open rows −24.1%, tight-budget compile tokens −59.5% — measured, not modeled. When a judge or dataset gates a number, that's stated as a named residual, not rounded up.