News · 2026-07-31

Metis puts an agent's memory inside the model instead of in a database beside it

A team led by researchers from MemTensor and several Chinese universities released Metis on July 29, the first prototype of what they call a memory foundation model: a language model with a persistent memory state built into the backbone, updated after every interaction by a single forward pass with no gradients involved, and read through a dedicated memory-attention path. The paper drew 241 upvotes on Hugging Face's daily papers feed, among the highest of the day, and the checkpoints are public.

Key facts

• Metis maintains a memory state inside the model; updates require only a forward pass and no gradient computation.
• All learned model weights stay frozen at inference; only the memory state changes.
• Released July 29, 2026, with project and model checkpoints public at MemTensor/Metis.
• Primary source: Metis: Memory Foundation Model.

Almost every AI system that claims to remember you does it the same way: memories live outside the model, in a database, and a retrieval step pastes the relevant ones back into the prompt before the model answers. That approach works, it is inspectable, and it is what agent memory means in practice today. It also has a hard ceiling - everything the model "remembers" has to fit through the context window, and something has to decide what is relevant before the model has thought about the question.

Metis moves the memory inside. The authors formalise native memory along two lines: a persistent, dynamically evolving state within the backbone, and native memory procedures - the model storing and using information through its own computation rather than through an external module. Their framing is that agent capabilities have steadily been absorbed into foundation models, giving rise to multimodal models and reasoning models, "however, agent memory is still primarily implemented through external modules, leaving the native memory capability largely unexplored."

The closest everyday analogy is the difference between a notebook and a skill. A notebook is external: you write things down, and later you find the page and read it back. A skill is internal: you drove a car this morning without retrieving a memory of driving lessons. Metis is an attempt at the second kind. Historical information gets compressed into a memory state, and the model reaches it through attention rather than through lookup.

The engineering constraint that makes this practical is that the update is gradient-free. Ordinarily, changing what a network holds means backpropagation - computing gradients and adjusting weights, which is far too slow and too destabilising to do after every conversation. Metis instead transforms its memory state through standard forward computation, the same kind of arithmetic the model already does to produce a token. That is what makes updating after each interaction affordable.

The paper is unusually forthcoming about where this breaks. Repeated updates weaken earlier stored facts, interference accumulates, and irrelevant accumulated state drags down general performance. That is the memory equivalent of a whiteboard nobody erases: at some point new writing goes on top of old writing and neither is legible. The authors publish this analysis themselves - "a detailed analysis of its strengths, limitations, and behaviors" - rather than leaving it for reviewers to find.

Getting a model to acquire this behaviour is its own problem, and it is where most of the engineering went. A model does not spontaneously learn to write useful things into a memory state and read them back later, because ordinary training gives it no reason to. The authors construct large-scale memory-specific training data and add several optimisation objectives during a mid-training phase, so that storing and retrieving become learned procedures rather than architectural hopes. The architecture supplies the place to put memories; the training supplies the habit of using it.

Metis landed the same week as several other approaches to the same word. Memory Decoder at Scale trains a separate 6.9-billion-parameter memory module offline and mixes its predictions with a frozen base model, reporting that pairing it with a 410-million-parameter Pythia lifts the average across 17 benchmarks from 29.86 to 37.34 - matching a 12-billion-parameter model with 39% fewer total parameters. MemHarness argues memory should be reconstructed rather than replayed verbatim. A filesystem-memory study tests the opposite bet, that memory should be plain markdown files an agent maintains itself.

None of these has been independently replicated - all are preprints days old, and the field is converging on the question rather than the answer. The useful thing Metis contributes is not a benchmark win but a sharp position: memory as a property of the model, not a service beside it.

Primary source, verified: read the paper → (arXiv 2607.26760)

Ground Truth. (2026, July 31). Metis puts an agent's memory inside the model instead of in a database beside it. Ground Truth. https://groundtruth.day/news/metis-gives-a-model-a-memory-that-updates-without-any-gradients.html

@misc{groundtruth:metis-gives-a-model-a-memory-that-updates-without-any-gradients,
  title  = {Metis puts an agent's memory inside the model instead of in a database beside it},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/metis-gives-a-model-a-memory-that-updates-without-any-gradients.html}
}

Topics: agent-memory · architecture · research · long-context · open-weights