Metis: Memory Foundation Model

Metis introduces a prototype of memory foundation models, where a persistent and dynamically evolving memory state is integrated into the model backbone rather than implemented as a separate agent module. Its architecture compresses historical information into native memory and retrieves it through memory attention. The authors construct large-scale memory-specific training data and use multiple mid-training objectives to learn memory procedures. During inference, model weights remain frozen, while memory states are autonomously updated through standard forward computation. Online maintenance is gradient-free and requires only a forward pass. The project and model checkpoints are released.