State-Level Correctness: The Case Against Databases for Agent Memory

The paper demonstrates that database abstractions fundamentally fail for agent memory because correctness must exist at the state level, not the record level—a critical insight for building long-running agents. The proposed GEM model replaces traditional CRUD operations with state-aware alternatives (ingestion, revision, forgetting, and coupled retrieval), directly addressing four failure modes that plague current systems: unregulated growth, missing semantic revision, capacity-driven forgetting, and read-only retrieval. This reframes how developers should architect agent memory: not around database schemas, but around how memory state evolves and transforms over time.