AI memory stores fragments of past sessions and re-reads them as though encountering them for the first time.

Only 38% of information stored by ChatGPT's memory met basic standards for security, accuracy, and relevance.

Adults can compare a vague profile against a stable sense of self. Children are still building one.

Students looked up as psychologist Bertram Forer entered the room. He handed his introductory psychology class a personalized personality assessment generated from a diagnostic test they had taken earlier in the course. He asked them to rate its accuracy from zero to 5. The average rating was 4.26/5.

What Forer hadn’t told his students was that this test was a trick. In fact, he had assembled the profile from a newsstand horoscope a day earlier and had given all of his students the exact same profile. The statements were written to feel specific while staying true to almost anyone. "You have a tendency to be critical of yourself." "You prefer a certain amount of change and variety." Each student read these sentences as a precise description of who they were. The effect Forer identified, later called the Barnum Effect, has been one of the most replicated findings in personality psychology. People accept vague, general statements as applicable to them when they believe the statements were written for them specifically.

I’ve been thinking about this experiment a lot since reading a study by Sadat and colleagues that analyzed what ChatGPT's memory feature actually stores about its users. As you may assume at this point, it’s akin to a circus trick.

What AI Memory Actually Does

If you have enabled the AI memory as a feature on your LLM you may wonder how the model is able to perform recall........

© Psychology Today