Tsinghua benchmark finds AI agents rarely forget personal details cleanly

Tsinghua University researchers introduced BEMULA, a benchmark that tests whether AI agents with long-term memory actually forget a specific personal detail when asked.

Across six memory systems and three backbone models, no tested setup passed more than 1.5% of forget requests end to end, according to a report on the paper. After a forget request, systems blocked at most 35.7% of simple extraction prompts, such as role-play or yes-or-no guesses. The report says the assistants either kept the detail or deleted too much, so a confirmation message does not mean the data is gone.