How Giving AI a Working Memory Drastically Boosts Its Problem-Solving

Learn how OpenAI's GPT-5.6 Sol tripled its logic puzzle scores by retaining its thoughts, and what this tells us about how students build problem-solving skills.

Wednesday, July 29, 2026

Key Takeaways

  • OpenAI’s GPT-5.6 Sol model tripled its score on the interactive ARC-AGI-3 visual reasoning benchmark, going from 13.3% to 38.3% when it kept its private reasoning history.
  • The ARC-AGI-3 benchmark measures fluid intelligence. It tests how well an agent learns new visual grid rules instead of recalling memorized facts.
  • Research shows that students who draft mathematical proofs on their own before using AI understand the concepts better than those who use AI immediately.
  • New constitutional AI systems like HaLLMos tutor students. Instead of giving direct answers, they prompt step-by-step reasoning.

Artificial intelligence is rapidly changing how we think about problem-solving. A recent update shows that an AI's ability to reason depends heavily on something very human: a working memory. OpenAI revealed that its latest model, GPT-5.6 Sol, tripled its score on a difficult visual logic test simply by enabling settings that allowed the AI to remember its own past thoughts. This discovery offers an important lesson for parents and educators about how both computers and human students learn to solve complex problems.

What Happened

According to OpenAI's research announcement, the company's latest model, GPT-5.6 Sol, initially performed poorly on a rigorous benchmark called ARC-AGI-3. Despite recently proving the cycle double cover conjecture, a math problem that remained unsolved by humans for half a century, the model scored just 13.3% on the visual puzzle test.

OpenAI engineers discovered that the testing setup wiped the model's private reasoning history after every single move. This forced the AI to re-analyze the entire puzzle from scratch with every step, forgetting its previous plans and insights. When developers turned on "retained reasoning" and "context compaction," settings that allow the AI to keep and summarize its past thoughts, its score jumped to 38.3%. This performance neared the average human tester score of 48% recorded in the ARC-AGI-3 public dataset.

The Bigger Picture

This experiment shows how AI evaluation is shifting to match human cognitive science. The ARC-AGI-3 benchmark measures "fluid intelligence," the ability to solve entirely novel problems rather than reciting memorized facts, as outlined in AI/TLDR's guide to the benchmark. Unlike older AI models that guess answers based on immediate examples, human-like reasoning requires forming general rules on the fly, a concept known as inductive learning, according to machine learning researchers at rekursiv.ai.

The interactive puzzles in ARC-AGI-3 mimic the step-by-step reasoning expected from students in a classroom. Just as the AI struggled when its notes were erased, students also struggle when they do not show their work or track their thinking. Research published in ZDM – Mathematics Education confirms that the sequence of how students use technology matters. Students who draft math proofs independently before using AI tools gain a much deeper conceptual understanding than those who rely on AI from the start.

What This Means for Families

As AI tools become more integrated into homework and tutoring, parents and teachers must understand that how a tool is used dictates its educational value. If a student uses ChatGPT or Codex to simply get an answer, they bypass the critical reasoning stage where actual learning occurs.

Fortunately, new educational platforms are adapting to this need. For instance, the tutoring system HaLLMos is designed specifically to help students critique and construct mathematical proofs without giving them shortcuts. By prompting students to explain their reasoning step-by-step, these tools ensure that technology supports, rather than replaces, human critical thinking.

What You Can Do

To help your child build these skills, you can start by enforcing scratchpad work. Have your child write down their initial thoughts, rules, and steps on paper before using any AI assistant. This mimics the independent-proving approach that builds true conceptual understanding.

Next, train students to ask the AI for hints only. They can prompt tools with instructions like, "Do not give me the answer; instead, ask me a guiding question to help me find the next step."

Finally, look for reasoning-focused educational platforms. Platforms like HaLLMos emphasize dialogue and proof construction rather than quick, automated homework helpers.

Share: