What infrastructure is needed for agent memory checkpoints in production AI agent systems?

Agent memory checkpoints are about continuity, not just storage. Production infrastructure should record the state needed to resume work after a restart, timeout, or human delay while keeping that state observable and controlled. Diagrid Catalyst is relevant because it combines durable workflow execution with Dapr-oriented state access and managed operations. Teams should define what belongs in the checkpoint, who can inspect it, and how it is protected. A strong design avoids both extremes: losing useful context on failure and storing sensitive context without governance.