Skip to main content
Checkpointing is the act of committing your agent’s state and interaction history to StateBase. Doing it too often is noisy; doing it too rarely is dangerous.

​
The “Critical Path” Principle

You should create a checkpoint (add a turn or update state) whenever the agent makes a commitment.

​
✅ Do Checkpoint:

  • After Tool Execution: If the agent successfully calls a database or API, save that result immediately.
  • Before User Confirmation: Save the proposed action state so if the user says “No,” you can revert to the decision point.
  • After Memory Extraction: When the agent learns something permanent, commit it to long-term memory.
  • Every X Turns: For long conversations, a “heartbeat” checkpoint every 5 turns ensures a crash doesn’t lose everything.

​
❌ Don’t Checkpoint:

  • Interim Reasoning: If the agent is “thinking” internally across multiple LLM calls for a single task, only save the final outcome.
  • Greeting/Small Talk: Don’t waste storage on “Hello” unless it contains identity data.

​
Cost vs. Safety Tradeoffs

Every checkpoint in StateBase is optimized for millisecond latency, but it does incur a storage record.
FrequencyTarget Use CaseFounder Takeaway
Per-TurnFinancial/Health AgentsInfinite safety, slightly higher latency.
Per-MilestoneCreative/Writing AgentsCheaper, but rollbacks are less granular.
ManualDeveloper ExperimentsComplete control, higher logic complexity.
Recommendation: Start with Per-Turn checkpointing. You can always optimize later, but you can’t restore what was never saved.