Estudos
24 de ago. de 2026
Studies — when one vector was not enough: the memory that learned to speak two languages
Yurumi stored its memory with a single embedding model trained in Portuguese. It worked — until the first question in English. The fix was not swapping the model; it was running two of them side by side and fusing the rankings with RRF. After a full LoCoMo run, recall went from 0.139 to 0.1614.
Continuar lendo →