Looking at several agent memory works, essentially redoing traditional KG methods—entity disambiguation, coreference resolution, multi-hop reasoning, temporal KG... Haha, I remember about a year ago there was a vague argument that LLMs had overturned traditional NLP research, but now it seems they have to bring them back. Connectionism and symbol systems are always intertwined like this, each influencing the other, ebbing and flowing, never ceasing...