Search papers, labs, and topics across Lattice.
This paper introduces Metis, a novel memory foundation model that integrates native memory capabilities directly into the architecture of foundation models, rather than relying on external memory modules. By formalizing native memory as both a persistent state and autonomous memory procedures, Metis enables efficient information storage and retrieval through memory attention, optimizing end-to-end performance. Experimental results demonstrate that Metis not only enhances the model's memory capabilities but also provides insights into its architectural strengths and limitations, paving the way for future advancements in memory-augmented AI systems.
Native memory in foundation models can significantly enhance efficiency and performance, as shown by the innovative Metis architecture.
Recent advances in AI agents have increasingly internalized native capabilities into their underlying foundation models, giving rise to multimodal foundation models and large reasoning models. However, agent memory is still primarily implemented through external modules, leaving the native memory capability largely unexplored. In this paper, we take a first step toward this direction by introducing memory foundation models, which empower foundation models with native memory capabilities. We formalize native memory from two perspectives: a persistent and dynamically evolving memory state within the backbone, and native memory procedures that autonomously store and utilize information through model computation. We show that native memory offers advantages in architecture, end-to-end optimization, and efficiency. Based on this formulation, we propose Metis, the first prototype of memory foundation models. Metis introduces a new architecture that equips a foundation model with a native memory state, allowing historical information to be compressed into the model and accessed through memory attention. We construct large-scale memory-specific training data and introduce multiple optimization objectives to acquire these native memory procedures through mid-training. The online memory maintenance of Metis is gradient-free, and the memory update requires only a forward pass. At inference time, all learned model weights remain frozen, while the native memory states are autonomously transformed through standard forward computation. Through extensive experiments, we show that Metis exhibits native memory capabilities and further provide a detailed analysis of its strengths, limitations, and behaviors. To facilitate future research on memory foundation models, we release our project and model checkpoints.
