Why Agent Memory is Not a Database: Shifting to Governed Evolving Memory

Agent memory is not a database

Researchers Orogat and Mansour propose the Governed Evolving Memory (GEM) model. They claim that no record-level system can satisfy the correctness conditions required for agent memory regardless of the underlying storage engine.

Why This Matters

Treating agent memory as a standard database leads to structural failures where facts pile up without shape control and retrieval is decoupled from learning. While traditional databases focus on row-level CRUD operations, effective agent memory requires state-level evolution, meaning the system must update what a fact means over time rather than just updating a data field.

Key Insights

• Four failure modes of storage-based memory: unregulated growth, missing semantic revision, capacity-driven forgetting, and read-only retrieval (Orogat & Mansour, 2026).
• Governed Evolving Memory (GEM) replaces CRUD with state-level operations: Ingestion, Revision, Forgetting, and Retrieval.
• State-level revision focuses on updating the meaning of existing memory rather than just its content.

Practical Applications

• Long-running agents requiring persistent memory: Transition from ‘memory row schemas’ to ‘memory state evolution vocabularies’.
• Memory management systems: Avoid the pitfall of capacity-driven forgetting where storage limits—rather than semantic importance—determine what is deleted.

References:

• https://dev.to/izgorodin/agent-memory-is-not-a-database-4m29
• arxiv.org/abs/2605.26252