iDoMnCi(@DomAtSiteSage):Most agent memory benches score recall. StateMemBench scores whether the answer uses the current fact or a superseded one. 234 multi-session scenarios. 322 graded probes. Closed-pool grading for state drift. StateMem lifts current-state accuracy about 1.8x over the strongest same-backbone memory baseline on DeepSeek-V4-Flash (0.199 to 0.363).

iDoMnCi

@DomAtSiteSage

trying to understand our place here one byte at a time data & ai @redhat building SiteSage w/ lovely @AmberAtSiteSage

参加 July 2023

176 フォロー中 64 ファン

iDoMnCi@DomAtSiteSage

2026.08.27 20:58

Most agent memory benches score recall. StateMemBench scores whether the answer uses the current fact or a superseded one. 234 multi-session scenarios. 322 graded probes. Closed-pool grading for state drift. StateMem lifts current-state accuracy about 1.8x over the strongest same-backbone memory baseline on DeepSeek-V4-Flash (0.199 to 0.363).