Open-Source AI Just Passed a 196-Step Memory Marathon

Ever ask an AI to do something long, like plan a vacation or write a report, only to have it forget what it said earlier? That's a familiar AI weakness. When each step depends on the last, even tiny mistakes pile up. One wrong turn early in a 20-step process can ruin everything by the end. This paper attacks that exact problem.

The test was a marathon of memory: they asked the AI to compute an MD5 hash, which is a kind of digital fingerprint. MD5 uses 64 rounds of math on 32-bit numbers, a chore with 196 connected steps. Think of it as walking 196 blocks, where each new street depends on remembering the last. Remarkably, gpt-oss-120b, an open-source AI less powerful than some paid rivals, carried all that information correctly through most runs.

Two tricks made the difference, and neither required retraining the AI. First, the AI was allowed to keep its own notes and reasoning in its 'mind' at every step, rather than starting fresh. Second, a second AI reviewed and 'voted' on the arithmetic to catch slip-ups. That's like having a quick-thinking assistant who writes everything down and double-checks the math before moving on.

So why should you care? Because this is one of the clearest signs yet that AI can handle long, multi-step jobs that people rely on daily: booking refunds, managing timelines, even helping with research. The catch, of course, is it isn't perfect. It still stumbles on math sometimes, and this test doesn't cover messy real-world goals. But knowing where AI fails matters. The more separate a memory test is from other AI skills, the easier it is to fix the weak spots and make agents you can actually trust.