OpenAI ran the same model on the same AI benchmark in two different setups.

With the official setup, GPT-5.6 Sol scored 13.3%.

After OpenAI changed how the setup handled memory, the score reached 38.3%. The model also used roughly six times fewer output tokens.

Nothing about the model changed. It was not retrained, and nobody found a magic prompt. The second setup allowed it to remember its own work.

I am Alex, welcome to ShortCu8 by Innov8.

Lets Dive Deep 🐰

When an AI agent gets worse during a long task, inspect its working memory before replacing the model.

A long agent run is produced by the model and the software carrying its plans, earlier actions, tool results and remaining context.

If the agent keeps rediscovering the same facts or reviving ideas it already rejected, the brain may be fine. Its notebook may be missing pages.

ARC-AGI-3 gives an AI agent unfamiliar 2D games without written instructions.

The agent must press buttons, observe what changes, infer the rules and improve through experience. The score rewards completing levels with fewer actions relative to human players.

This makes memory important. A useful discovery on move 20 should still affect move 80.

In the official generic harness, two things got in the way.

First, the model's private reasoning was discarded after every game action. It could see past moves and brief notes, but it lost the plans and deductions behind them.

Second, the harness used rolling truncation. Once the history grew too large, it removed the oldest actions.

The model had to keep reconstructing the game while parts of its experience disappeared.

That is like debugging an app with a terminal history that shows every command but deletes why you ran it and what each failed test proved.

OpenAI rebuilt the test using the Responses API and enabled:

With those settings, GPT-5.6 Sol spent less time working out the game again. It kept a more coherent strategy across the run.

On the public task set, its score rose from 13.3% to 38.3%.

The experiment does not prove that every agent becomes three times better with these settings. OpenAI ran the comparison on its own model and implementation.

It proves something narrower and more useful: a benchmark score can measure the harness as much as the model inside it.

ARC used a generic harness deliberately. A common setup can make comparisons cleaner and expose what a model cannot do without product-specific help.

OpenAI tested a different question: how does the model perform with the memory features used in ChatGPT and Codex?

Both results are useful, but they describe different systems.

So when a leaderboard says one model beats another, check the harness, context limit, reasoning retention, tools and compaction method. The name beside the score is only part of what was tested.

Now go and built something great!

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.

Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.

Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.

Try Wispr Flow free

#AD2

Opening entities in every country can be slow, expensive, and hard to scale.

That's why more companies are using EOR to hire globally faster.

See how Oyster helps teams hire, pay, and support talent in 180+ countries while staying compliant along the way.

Get all the answers