Sakana Continuous Memory Machine Splits Short and Long AI Memory

The short version

Sakana AI shared a new AI design on 8 October 2026. It is called Continuous Memory Machine. The idea is to give the model two memory banks instead of one. One bank holds fresh notes. The other holds older facts. The paper is on arXiv. It is a preprint and is not yet peer reviewed.

What the researchers did

Ciaran Regan, Kai Arulkumaran, Luke Darlow, Stefania Druga, Sebastian Risi, and Llion Jones work at Sakana AI. They posted the paper on the arXiv site on 8 October 2026. They also shared code on GitHub under SakanaAI continuous memory machines.

Continuous Memory Machine uses two matrix shaped memory stores. Short term memory tracks recent neuron activity. Long term memory keeps facts for later steps. A Transformer module updates both stores.

Photo: Pexels

What they found

The team tests the model on algorithmic tasks, in context learning tasks, and loop style reasoning tasks. They report that Continuous Memory Machine beats listed baselines.

Why it matters for you

Agent teams need memory that keeps fresh notes apart from long term facts. Split stores may reduce summary loss in long chats.

What to keep in mind

This is a preprint, not yet peer reviewed. Compute cost rises when two stores update each step.

Key takeaways

• Two memory banks replace one crowded loop vector.
• Transformer links both stores with two way updates.
• Code is public for repro runs.

Sources

• Primary — arXiv, Continuous Memory Machines (8 October 2026). Defines the model and lists benchmark claims.