memory voting sounds interesting but does it work? i tried having agents mark useful chunks once, they just marked everything as helpful. accuracy went to shit