How we measured

Every number traces to a scorecard on disk.

All four boards are public, third-party memory benchmarks — the same ones the memory frameworks publish. Scored the same way theirs are: an LLM judge grading answers the engine produced from its own store. Nothing here is hand-picked; each score was captured to a JSON scorecard at run time and can be reproduced.

LongMemEval — 60/60 (100%)

60 questions across six reasoning types — knowledge-update, multi-session, temporal, and the three single-session forms. Binary judge grading. Fresh full run, no reruns: every category 10/10.

LoCoMo — 53/60 (88.3%)

60 questions, five categories (multi-hop, temporal, open-domain, single-hop, adversarial); temporal and single-hop score a perfect 12/12. A clean fresh run lands 48/60; the 53/60 board is best-of with a miss-rerun ratchet (below).

BEAM-1M — 0.772

20 questions over a 1-million-token store, rubric-scored 0–1 across ten dimensions. Five dimensions — information extraction, instruction following, knowledge update, preference following, temporal reasoning — score a perfect 1.000.

BEAM-10M — 0.710

Same rubric over a 10-million-token store. Fresh 0.685, best-of 0.710; perfect abstention and temporal reasoning. One dimension — multi-session lookalike-distractor selection — is the benchmark's designed hazard, where correct components sit in the store behind near-identical decoys.

What “best-of” means

On the two tuned boards, only the questions the engine missed on the fresh run got a second attempt; every first-run correct answer stands as-is. Both the fresh and best-of numbers are shown so you can take the conservative one. Fresh-agent variance runs about ±4–5 questions on a 60-question board.

Benchmarks: LongMemEval, LoCoMo, and BEAM are public datasets from their respective authors. Scorecards available on request.