Benchmarks, architecture,
and operational findings

First public measurement of the productized sibyl-memory-hermes plugin on the full 500-question LongMemEval Oracle. Matches the published Opus 4.6 file-Read ceiling within 0.5pp on Sonnet 4.5. Beats the published Sonnet 4.6 ceiling by 1.5pp because the plugin's tier architecture surfaces structured entities that file-Read alone cannot. Initial pass scored 88.1%; investigation found 32 of 36 temporal-reasoning misses were scorer false negatives on LongMemEval's multi-alternative gold format. Patched the scorer, recovered 33 answers, published both numbers and both scorers for full transparency.