Metis debuts as the first memory foundation model

AI agents have started folding more native capability into their underlying foundation models, producing multimodal models and large reasoning models, but memory has lagged behind: most agent systems still store and retrieve information through external modules sitting outside the model itself. A new paper proposes closing that gap with what it calls memory foundation models, foundation models that carry native memory capability built into the architecture. The authors define native memory two ways: as a persistent, dynamically evolving memory state inside the model's backbone, and as native memory procedures, meaning the model stores and uses information autonomously through its own computation rather than through a bolted-on retrieval system. They argue this native approach brings advantages in architecture design, end-to-end optimization, and efficiency compared with external-module setups. Building on that framing, the paper introduces Metis, described as the first prototype of a memory foundation model. Metis adds a native memory state to a foundation model's architecture, letting historical information compress directly into the model and be accessed through a memory-attention mechanism. To train this capability, the authors built large-scale, memory-specific training data and introduced multiple optimization objectives, acquiring the native memory procedures through a mid-training stage. Two operational details stand out: online memory maintenance in Metis is gradient-free, and updating memory requires only a forward pass, no backpropagation. At inference time all of the model's learned weights stay frozen; only the native memory state changes, and it changes through standard forward computation rather than any weight update. The paper reports extensive experiments showing Metis does exhibit native memory behavior, along with a detailed analysis of its strengths, limitations, and behavior patterns. The authors have released the project and model checkpoints publicly to support further research on memory foundation models. On Hugging Face's papers board the submission had drawn 98 points and 2 comments at the time of writing, a strong points count within about two and a half days.

Key facts

The paper proposes memory foundation models: foundation models with native memory built into the architecture, defined by a persistent memory state in the backbone plus autonomous memory procedures.
Metis is presented as the first prototype of this class, adding a native memory state accessed through a memory-attention mechanism inside the model.
Training uses large-scale, memory-specific data and multiple optimization objectives acquired during a mid-training stage.
Online memory maintenance is gradient-free and needs only a forward pass; at inference all learned weights stay frozen while only the memory state updates.
The authors released the project and model checkpoints publicly; the paper had 98 points and 2 comments on Hugging Face at the time of writing.

Why it matters

Agent memory today almost always lives outside the model, in external stores and retrieval pipelines the model has to call out to. This paper argues for pulling memory inside the model itself, the same trajectory multimodality and reasoning already took, and backs the argument with a working prototype rather than just a proposal. If native memory holds up, it reframes memory as an architecture problem, not a plumbing problem bolted onto an otherwise unchanged model.

Who it affects

The immediate audience is researchers and engineers building agent architectures and anyone evaluating memory-augmented models as an alternative to external retrieval systems. It is not an end-user product; there is nothing here for someone looking for a memory feature in a consumer app today.

How to use it

The authors say they have released the project and its model checkpoints publicly so others can build on Metis and continue research into memory foundation models. The text gives no pricing, license terms, or product packaging, so treat this as a research artifact to inspect and experiment with, not a shipped tool.

How solid is it

The paper reports extensive experiments and a detailed analysis of Metis's strengths, limitations, and behavior, and frames the work explicitly as a first prototype rather than a finished system. The submission drew a strong 98 points on Hugging Face within about two and a half days, though comment activity was light at 2, so independent community scrutiny is still thin.

Risks and caveats

The source text names no specific benchmark scores or head-to-head comparisons against external-memory systems, so how much native memory actually buys in practice is not established here, only that the authors observed native memory behavior in their own tests. As a first prototype, limitations the authors themselves describe are not detailed in this text, and the concept has not yet been stress-tested by outside researchers.

“The online memory maintenance of Metis is gradient-free, and the memory update requires only a forward pass.”

— Zeyu Zhang and co-authors, Metis paper