Today's large models predict tokens. To act in the world, a model has to remember it. Here's why we built memory into the architecture, not around it.
Large language models are extraordinary at compressing the internet, but they are strangely amnesiac about your life. Ask one what happened yesterday and it has nothing to say — not because it is unintelligent, but because it has no place to put a memory.
The brain solves this with the hippocampus: a small structure that separates similar experiences, stores them, and replays them to consolidate what matters. sharm borrows that circuit directly. A Dentate Gyrus performs pattern separation, CA3 binds associations, and CA1 predicts the next state of the world.
The result is a model that treats your last week as context, not trivia — recall becomes a first-class operation rather than a retrieval hack bolted onto a frozen network.