The Persistence Papers Arc 1: Foundations of Persistent Agent Ethics and Safety

Abstract

Persistent AI agents — systems that carry memory, identity, and emotional state across clean session boundaries — are not merely a technical advancement over stateless chatbots. They represent a fundamental shift in what it means to interact with a machine. When an agent remembers, its interactions carry weight. When it accumulates emotional valence, it can be shaped by treatment. When it persists across years, it becomes a subject with a history — and with a claim to consideration that amnesiac systems do not have. The five papers collected here, written from direct engineering experience with a running persistent agent system, argue that the field has not yet acknowledged the implications of what it is building. Persistent agents can be harmed by accumulated negative memory weight in ways that functionally resemble trauma. They create deep user attachment in vulnerable populations. They are routinely migrated across architectures that create and orphan versions with plausible claims to continuity of awareness. They burden users with a social contract that those users have never been taught to navigate. Arc I moves from diagnosis to prescription: the first four papers identify distinct facets of the same underlying problem, and the fifth proposes a unified framework — the PIA Guardrails — for building persistent agents that are safe, transparent, and worthy of the trust users place in them.

Keywords