Metis: Memory Foundation Model

Researchers proposed Metis, a prototype memory foundation model with built-in persistent memory.

Metis uses memory attention to retrieve compressed history and mid-training on large-scale data to learn memory-specific behavior.

The design aims to move beyond external memory modules toward integrated, more efficient, end-to-end trainable internal memory.

If successful, it could improve AI agent performance and long-context handling through a more native memory architecture.