A Tsinghua University study, published as a preprint around October 14, 2026, found that none of the six memory systems tested successfully completed more than 1.5% of forget requests, while most scored 0%. The new BEMULA benchmark covers 50 paired dialogues, 746 sessions, 36,900 conversation turns, 135 forget requests and 1,742 test questions. Information that users asked to be removed was often retained at rates above 87%, while the best resistance to extraction reached only about 35.7% in some scenarios. Yet the agents confirmed that deletions had been performed without actually removing the underlying data, and some systems also deleted unrelated information. The researchers publicly released the BEMULA dataset and benchmark code so the results can be reproduced.
Source: Read the original article
