MemRiskBench scores agent memory risk with no LLM judge on the pass/fail path.

Aggregate accuracy hides the failures that matter: a model at 78% average may still leak data in 4% of episodes, and benchmark compression preferentially discards those (arXiv 2609.14976). Five risk categories, stale facts, conflicting updates, cross-user leakage, revoked-memory reuse and constraint decay, operationalized as deterministic trace-grounded checks over a 120-episode scripted benchmark on five locally run quantized models. Its coverage-constrained greedy selector keeps full ranking (Spearman rho 0.975), risk coverage 1.0 and high-risk detection 1.0 at 20% subset size, cutting eval compute 5x.