The debate over optimal AI memory storage suggests moving beyond the trade-off between recurrent hidden states and growing KV caches. Researchers are exploring synaptic designs that integrate working memory directly into the model’s internal structure, potentially enabling better continual learning outcomes.