Memory that keeps its receipts
Extracted facts answered 84.4% of LongMemEval-S. Linking each fact back to the conversation it came from raised that to 89.6%, for 20% more storage.
Tagged
evaluation
← All postsHow we took LongMemEval from 80.0 to 94.2 without touching retrieval
LongMemEval-S climbed from 80.0 to 94.2 while retrieval recall at 10 stayed at 0.99. The gains came from evidence budget, reader generation, and instructions.
We retrieved the memory, then dropped the answer
A 200-question conflict run scored 0.800 at the parent-Memory boundary and 0.635 at the evidence boundary. The gap was in what reached the reader.