Evidence · DiggingBeagle record
FEB-30: Beyond Downstream QA for Agent Memory
Original benchmark measuring stored evolving facts independently from final question answering.
- Source role
- primary disclosure
Evidence record
Original benchmark measuring stored evolving facts independently from final question answering.
Claim-level citations (1)
- supportsCorrect agent answers concealed incorrect persistent memory states: FEB-30 reports an average sixteen percentage-point gap between internal fact-state correctness and final question answering, including memory operation classification accuracy of only twenty percent in the tested system.
Abstract; 6 Experiments; Findings 1 and 3