​

Cost vs. Safety

​

The Tradeoff

​

What Actually Costs Money

1. State storage — every checkpoint is a version of your state
2. Turn storage — every logged interaction
3. Memory embeddings — semantic search indexing
4. API calls — retrieving large contexts

• Store only state, not raw transcripts. If you don’t need full transcripts for auditing, persist state and a summary.
• Sliding-window turns. get_context(turn_limit=10) reads 10 turns but state history stays intact.
• Summarize old turns into memory instead of keeping them verbatim.
• TTL on caches. Don’t checkpoint throwaway data (greetings, transient lookups).

​

The 80/20 Rule

• Before calling a paid/external API
• Before executing a user-visible action
• After a state mutation that later steps depend on
• When the user confirms or changes a decision

def commit_point(session_id, state, reasoning):
    sb.sessions.update_state(
        session_id=session_id,
        state=state,
        reasoning=reasoning,
    )

​

Budgeting Model

Checkpoint cost   ~ $0.000001 per 1KB state (storage)
Turn cost         ~ $0.0001 per turn (retrieval + storage)
Memory embedding  ~ $0.0001 per memory item

An agent doing 1,000 turns/day, per-turn checkpointing:
  ~ $0.10/day  in state
  ~ $0.10/day  in turns
  = ~ $6/month  per active agent

​

Recommendation

​

Next Steps

• When to Checkpoint: the critical-path principle
• Rollback Strategies: how to recover
• Reliability Guarantees: what happens when things go wrong