How do you know your agent's memory layer is helping, not hurting?

I work on long-term memory for agents, and the honest problem I keep hitting: everyone benchmarks the model, almost nobody benchmarks the memory layer itself. A bad memory layer can actively make a good model worse, stale facts retrieved with high confidence, contradictions never surfaced. The vendor answer is usually "trust our evals." I published an open harness instead, MIT, that reruns 500 LongMemEval-style questions against any local stack so the numbers are checkable, not claimed: https://github.com/yadu9989/memstrata-lme500-harness

Curious how others here are evaluating theirs. Do you trust the memory layer, or treat it as guilty until proven innocent?