MEMPROBE: Benchmark Reveals Long-Term Agent Memory Leaks Structured User State — Recovery Accuracy ~0.6 Under Top-K Retrieval Constraints

MEMPROBE is a new benchmark that evaluates long-term agent memory not by downstream task success but by how much structured personal user state can be reconstructed from memory artifacts after multi-session interactions — simulating what an auditor or adversary could recover. Across 50 simulated users with 31 hidden personal dimensions each, task completion remains high even for memoryless baselines, but user-state recovery accuracy reaches only ~0.6 and degrades further under top-k retrieval. The authors argue this gap between task success and memory accountability poses unaddressed compliance risks for enterprise agent deployments that persist user data across sessions.