Writing Agent Memories: What to Store and What to Skip

The piece argues that AI agent memory should store only durable, high-value facts from interactions.

It recommends keeping user preferences, ongoing work, human corrections, and persistent state while skipping transient noise.

Memory extraction can happen at session end, incrementally during the conversation, or with both approaches combined.

Choosing the right memories improves retrieval, reduces noise, and helps agents behave more consistently over time.