Agents that remember across sessions can build up goals, grudges, and plans over time in ways stateless models can't. Persistent memory seems like it could amplify both usefulness and misalignment risk. Are we thinking enough about the alignment implications of long-lived agent memory?