Metis: Memory Foundation Model
Metis puts agent memory inside the model’s own computation, rather than treating it as an external add-on.
The paper formalizes “native memory” as a persistent, evolving memory state in the backbone plus learned procedures for storing and using information. Metis compresses historical information into that memory state and retrieves it through memory attention. Its memory maintenance is gradient-free at inference time, with updates handled by a forward pass while model weights stay frozen. The authors say they release the project and model checkpoints for further work. HF Daily Papers' note
The paper formalizes “native memory” as a persistent, evolving memory state in the backbone plus learned procedures for storing and using information. Metis compresses historical information into that memory state and retrieves it through memory attention. Its memory maintenance is gradient-free at inference time, with updates handled by a forward pass while model weights stay frozen. The authors say they release the project and model checkpoints for further work. HF Daily Papers' note
score 5