ロスを防ぐ三層の検索 ★ ライブ
アクティブコンテキストのファクト圧縮により、有用な情報の約100%を保持できます。一方、一般的な圧縮では約40%に留まります。長期的なメモリ検索は別の分野であり、その性能は公開ベンチマークで評価されています。
A three-layer recall pipeline (L1+L2+L3) reconstructs context even after long gaps. Two distinct disciplines, stated honestly. (1) Compression of the ACTIVE context: replacing generic compaction with our useful-fact compression retains ~100% of facts on the same context window, versus ~40% with summary-style compaction — a data-quality measurement, reproducible benchmark being finalized on our public rig. (2) Retrieval from long-term MEMORY — across sources and time-validity windows — is a harder problem measured separately (LoCoMo, LongMemEval); numbers there are honest, not inflated — lower token cost, lower latency, nothing important lost.