Save policies

A save policy controls when messages created during a run are appended to the memory store.

Choose a policy

Policy	Behavior	Use when
'message'	Saves each completed user, assistant, and tool-result message.	Incremental durability matters most.
'turn'	Saves completed messages after each model-and-tool turn.	Product chat should preserve complete turns.
'run'	Saves only after a successful final response.	Partial runs must not enter future context.

'message' is the default. Choose the policy explicitly so the intended failure behavior is visible in agent configuration.

const agent = new AgentBuilder('support', model)
  .memory(memoryStore, { savePolicy: 'turn' })
  .build()

Failed runs

Messages already saved by 'message' or 'turn' may remain when a later step fails. With 'run', normal messages are written only after the run completes successfully.

A store may implement recordError(...) to retain failed-run diagnostics separately. Those records are operational data and should not be loaded as normal conversation context.

Match the policy to side effects

Save policy controls transcript persistence, not tool execution. A failed run may already have called a tool, so side-effect tools should remain idempotent or independently auditable.