The world is being quietly rearranged by people who write very long documents.


The title they went with A-MBER: Affective Memory Benchmark for Emotion Recognition Noisy translates that to

AI emotion recognition can't remember what it learned about you


Researchers built a benchmark to test whether AI assistants can track how your emotional state changes over time across multiple conversations. Current AI systems fail at this — they can read your mood in the moment but forget the pattern of how you felt yesterday, last week, or last month.
Right now, AI chatbots and assistants that claim to know you personally are actually starting fresh each conversation. This benchmark makes that failure visible and measurable. It shows that long-term emotional memory requires more than just having access to old conversations — the model has to learn to extract the subtle, changing patterns in how someone's mood evolves. This matters because if AI is going to function as a therapist, counselor, or long-term companion, it can't just pattern-match against the current moment. It has to understand that someone who was angry last week and sad yesterday might be withdrawn today for reasons connected to both.
Watch whether companies building long-term AI assistants (healthcare, mental health, personal AI) start publicly testing against affective memory benchmarks like this one, or continue ignoring the failure and marketing stateless systems as personalized.

If you insist
Read the original →