Most AI memory benchmarks test semantic recall. But coding agents don't really fail like that. They don't just "forget", they break their own earlier decisions while they're still in the code. So I built a benchmark for that…