Mnemosyne tops LongMemEval at 87.4%

This is a credible systems result, not just a flashy benchmark claim: the score looks driven by retrieval architecture, not brute-force model size.

• –Deterministic Spine indexing is the most interesting piece here; it should outperform embedding-only retrieval on exact facts, numbers, and preferences.
• –The 65.4% Multi-Session score is the real limitation, because cross-session accumulation and state drift are where memory systems usually break.
• –The flexible judge and cloud embeddings mean this is promising, but not a clean offline SOTA claim yet.
• –Flat RAM usage and low SSD I/O are the operational proof point: the architecture appears practical on consumer hardware.
• –The next meaningful benchmark is numeric state persistence across sessions, not more surface-level recall.