New AI Coding Tool Fails After 12 Hours, Forgets Its Own Work

A tech enthusiast recently put a new AI coding assistant called Strata through a grueling 12-hour test. The AI was supposed to write and test code non-stop. But after several hours, it began making bizarre mistakes. It forgot whether it had created a report, hallucinated that a file didn't exist, and even misread a folder name by dropping a single digit. These aren't small typos—they're signs the AI lost track of what it was doing.

The tester, who shared results on Reddit, isn't a beginner. He normally runs a similar AI model called Qwen3.8-Flash-Next and supervises it closely. He noted that the AI's memory seemed to degrade over time, especially after the conversation got very long. When the AI tried to recall earlier steps, it got confused and started making things up. This is a known problem called 'context window overflow'—the AI's short-term memory fills up and it starts forgetting or inventing details.

Why should you care? AI coding tools are being sold as productivity boosters that can work while you sleep. But this test shows they still need a human babysitter. If you're a manager thinking of replacing a junior developer with an AI, think again. The AI might write code for hours, but it could also undo its own work or introduce hidden bugs. For now, AI is a helper, not a replacement.

The catch: This was just one test with one tool. Other AI models might perform better. But the core issue—AI losing track over long tasks—affects all current systems. Until AI can reliably remember what it did, you should treat it like an eager intern: useful, but always double-checked.