Pulse Daily — Friday, October 9, 2026

Let's start with a story that has the AI research community buzzing this morning.

A system called Memento 3, built by researchers at University College London and Huawei's Noah's Arc Lab, posted a perfect score, 100.0, on Arc AGI 3, the hardest agentic reasoning benchmark in existence.

To understand why that matters, you need to know what Arc AGI 3 actually is.

It's an interactive benchmark that drops an AI agent into novel, abstract, turn-based game environments with absolutely zero instructions.

The agent has to figure out the rules, the goal, and the optimal strategy entirely on its own, just like a human would.

When the benchmark launched back in March, frontier AI systems, including the best models from OpenAI and Anthropic, scored below 1%.

Humans score 100%.

That gap was considered one of the clearest illustrations of what AI still cannot do.

And what makes Memento 3 so interesting is how it got there.

It did not achieve this by scaling up more compute or training on more data.

The paper, posted today, describes a system that keeps its underlying language model completely frozen.

It doesn't update the model weights at all.

Instead, the agent learns by rewriting a natural language rulebook stored as a markdown file alongside a compiled Python world model engine, updating both whenever its predictions fail.

It completed all 25 public Arc AGI 3 games using only 44% of the actions a human baseline required.

For comparison, Claude Opus 5, which had been one of the top performers on this benchmark, scored 40.7. Memento 3 scored 59 points higher.

This is a fundamentally different approach to intelligence, not memorizing patterns, but actively updating a model of the world.

And the ARC Prize Foundation has a $700,000 grand prize waiting for any system that can hit 100% on the official competition leaderboard.

That deadline is October 26th.

Today's paper is not yet that submission, but it is a very loud signal.