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...