Russian developers from "Agent Atom" at EMNLP 2026 showed a method for evaluating AI agents against Mem. As part of the test, the AI is implanted with changing facts - changing knowledge, available models. The study found that 55.6% of the answers recognized as quality standard indicators did not respond to contradictory information, while 38.8% did not respond to updated data. Agent Atom has open sourced the method so businesses can test their AI solutions. @cgevent.