Agent memory

When to use

Multi-step tasks with evolving user intent, preferences, or constraints.

Architecture sketch

Input → Scratchpad (ephemeral) → Tools
      → Profile (long-term) → Ledger (task)

Pitfalls

• Storing everything forever → privacy and drift
• Missing expiry windows and provenance

Checklist

• [ ] Define memory types and retention windows
• [ ] Source-of-truth + PII handling
• [ ] Summarization with confidence and citations

Good–Better–Best

• Good: ephemeral scratchpad
• Better: task ledger with rollups
• Best: governed profile with consent and expiry

Tiny eval snippet

def recall_rate(prompts, memory):
    return sum(int(intent in memory) for intent in prompts) / len(prompts)