A research team released PAST-Bench, a benchmark built to answer a question that most agent memory products simply assume the answer to: does giving an agent a persistent memory across sessions actually make it perform better later, or does it just make transcripts longer? Personal AI agents increasingly retain user preferences, past task outcomes, learned tool routines and other state between conversations, and vendors routinely market this as an improvement, but until now there hasn't been a controlled way to isolate whether the retained experience is doing real work. PAST-Bench runs each agent through the same ordered sequence of fresh-session tasks twice — once with retained memory switched on, once with it switched off — across 26 scenarios and 204 episodes spanning four categories: raw memory recall, reusing a previously learned procedure, gathering information across sessions, and correctly updating stale information once circumstances change. The interesting finding isn't just whether scores go up with memory enabled (they mostly do), but whether the improvement actually traces back through the intended pathway of saving something, retrieving it later, and using it correctly — as opposed to the agent getting lucky or the benchmark leaking information some other way. Testing seven base models across four agent frameworks, the authors found real but uneven gains: two agents can post the same headline improvement while one has clear evidence it followed the save-retrieve-use pathway and the other doesn't, which matters a lot if you're deciding whether to trust a memory system in production. The authors also built a reference implementation, Hermes+, with five targeted interventions across the memory lifecycle, and it improved both the size of the gain and how well that gain was evidenced, especially on tasks requiring an agent to recognize that old saved information is now wrong and needs replacing. For anyone building or buying an agentic system with memory as a feature, this gives a concrete methodology for checking whether that memory is actually earning its complexity budget rather than just accumulating write-only logs. Code and eval harness are on GitHub.