AI Learns Like a Human Brain—And It's Getting Smarter

Vision Transformers struggle to organize new information across memory timescales. A new paper proposes Hierarchical Hebbian Memory: three memory levels—rapid Working Memory, persistent Routed Episodic Memory, and slower Semantic Memory—with a learned controller managing memory contribution, routing, and consolidation. Tested on Omniglot 5-way 1-shot recognition and CORe50 continual object recognition, the model reaches 97.39% accuracy on Omniglot and 95.37% final accuracy on CORe50 when combined with experience replay, using a Swin-Tiny backbone. Its learned multi-bank retrieval also beats a single persistent bank and a no-memory baseline on a delayed-association task: 47.50% accuracy versus 24.17% and 25.00%, respectively. And after intervening distractors, Episodic Memory keeps roughly 0.96 cosine similarity with stored associations, while Working Memory drops to about 0.05. The findings show Hebbian association and learned memory routing can organize online visual experience across rapid, persistent, and consolidated memory timescales within Vision Transformers.