> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI Tripled AI Scores by Changing Two Settings
- URL: https://wire.fourthweb.ai/openai-tripled-ai-scores-by-changing-two-settings/
- Published: 2026-07-29T15:00:00.000Z
- Updated: 2026-07-30T04:32:40.000Z
- Description: OpenAI just accidentally proved that the path to better AI isn't smarter models—it's smarter memory management. OpenAI tripled GPT-5.6's scores on the ARC-AGI-3 benchmark by flipping two API switches: retaining reasoning traces and enabling context compaction
- Author: Travis Wright
- Tags: AI Agent Economy, AI Agents, Institutional Crypto, OpenAI

[**OpenAI**](https://wire.fourthweb.ai/tag/openai/) **just accidentally proved that the path to better AI isn't smarter models—it's smarter memory management.**

### The Summary

- [OpenAI tripled GPT-5.6's scores on the ARC-AGI-3 benchmark](https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores?ref=wire.fourthweb.ai) by flipping two API switches: retaining reasoning traces and enabling context compaction
- The breakthrough wasn't model architecture—it was letting the AI keep its work and compress what it learned
- This suggests the bottleneck in agent capability isn't raw intelligence, but working memory and information persistence

### The Signal

The ARC-AGI benchmark tests abstract reasoning—the kind of pattern recognition that makes humans smart beyond rote memorization. It's designed to be hard for current AI systems. GPT-5.6 was scoring respectably, but OpenAI researchers noticed something: the model kept solving the same sub-problems over and over, forgetting its own reasoning between tasks.

The fix was embarrassingly simple. First setting: retain reasoning traces across problem-solving steps instead of discarding them. Second setting: enable automatic compaction of that context so the model doesn't hit token limits. The result: a 3x score improvement and better computational efficiency. The AI wasn't getting smarter—it was finally allowed to remember what it had already figured out.

> "The bottleneck wasn't intelligence. It was institutional memory."

This matters because most production [AI agents](https://wire.fourthweb.ai/tag/ai-agents/) today are effectively amnesiac. They wake up fresh for every task, rebuild context from scratch, and forget everything when the session ends. That's fine for one-off queries, but catastrophic for complex workflows where the agent needs to build on its own discoveries.

The compaction piece is equally critical. Raw reasoning traces are verbose—every chain-of-thought step, every dead end explored, every correction made. Without compression, you hit context window limits fast. OpenAI's compaction takes those sprawling traces and distills them into dense, reusable knowledge. It's the difference between keeping a transcript of every meeting versus keeping well-organized notes.

What's not in the blog post but worth noting: this technique scales horizontally. An agent that retains and compacts its reasoning across tasks doesn't just get better at individual problems—it builds a library of reusable patterns. Solve invoice processing once, compress the approach, apply it to expense reports tomorrow. That's not AGI, but it's the architecture that makes agent deployment economical.

### The Implication

If you're building with AI agents, the lesson is clear: focus less on prompt engineering and more on memory architecture. The models are already capable enough for most business workflows. What they lack is the ability to learn from their own experience without human intervention each time.

Watch for memory-as-a-service tools to proliferate. The company that makes reasoning persistence and compaction dead simple for developers will own a big chunk of the agent stack. And if you're deploying agents at scale, ask your vendor how they handle cross-session learning. If the answer is "they don't," you're leaving 2x-3x performance on the table.

### Sources

[OpenAI Blog](https://openai.com/index/how-two-settings-tripled-our-arc-agi-3-scores?ref=wire.fourthweb.ai)