Agent Memory as State Trajectory, Not Storage
arXiv paper proposes Governed Evolving Memory (GEM) — a new data-management workload where agent memory correctness is defined by the entire state trajectory, not individual records.
Current systems fail in four ways: unregulated growth, missing semantic revision, capacity-driven forgetting, and read-only retrieval. GEM reframes long-term memory as a first-class problem requiring active governance, not passive storage.