Benchmarks, architecture,
and operational findings

An independent tester ran the same 500-company, 365-day business-memory benchmark across four engines: Sibyl, Hindsight, Mem0, and Mnemosyne. Sibyl retrieved every answer reading about 228 tokens per query; one competitor read 11,892 and still answered worse. The industry is optimizing for a bigger context window. The goal is the smallest correct one. Runner scripts, per-engine reports, and raw results included.

Independent testers ran Sibyl Memory 1:1 against Honcho on the same 42,000-record corpus, the same 250 questions, and Claude Sonnet 4.6 answering both. Sibyl answered 97.2% correct vs 85.6%, while retrieving 4.5x less context and costing 3.4x less. Plus 10x-scale retention (11,520 writes, 100% retained), an independent adversarial security pass, and the one relational-retrieval gap we are building a new graph-native elite tier to close. Every dataset and raw result is reproducible.

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.