Decision Log

​

In plain terms

The analogy: the meeting minutes that capture “we decided X because Y.” Months later, anyone can see why a call was made.

​

When to use it

• Application review — inspect a recorded decision and outcome alongside your host audit events and external service receipts. Model-written reasoning alone is not a complete audit record.
• Consistency — an approval or negotiation agent that should not contradict its own recent rulings.
• Learning from outcomes — log a decision, record whether it worked, and analyze quality over time.

memory: { storage, decisions: true }

​

When NOT to use it

• Simple chat agents with no consequential decisions — there’s nothing meaningful to log.
• If you only need raw conversation history, Sessions already capture it; the Decision Log is specifically for structured, queryable decision records.

​

Configuration

memory: { storage, decisions: true }                       // remind of last 5
memory: { storage, decisions: { maxContextDecisions: 10 } } // high-continuity agent

• Lower it (3) to save tokens or when past decisions aren’t relevant to the next one.
• Raise it (10) for agents that must avoid contradicting their recent rulings.
• It does not limit storage — only what’s surfaced in context.

​

What gets stored

{
  "decision": "Approved refund of $49.99 for ORD-7421",
  "reasoning": "Order was delayed 7 days, exceeding the 5-day SLA",
  "decisionType": "approval",
  "alternatives": ["Deny", "Partial refund"],
  "confidence": 0.9,
  "outcome": "Customer confirmed satisfaction",
  "outcomeQuality": "good"
}

​

Tools

​

Cross-references

• PII Guard — redaction for regulated data
• Memory Overview — the full memory lifecycle
• Reflection — agents that critique their own decisions