studies/memory-layers-for-agents/profiles/Profile__Delta-Mem

Memory as a dense rank-8 state matrix inside each attention head, updated token by token — the model never looks anything up.