Principles of Persistent Presence: Eight Design Principles for Long-Term AI Agent Architecture

Current AI memory systems treat memory as an information retrieval problem: store discrete facts, retrieve relevant records on demand. This technical note argues that for long-term AI agents, memory is not a storage module but the organ through which an AI remains present. We present eight design principles distilled from 18 months of building the Celestelin ecosystem (~130,000 lines of code across six architectures). These principles are organized in four layers — from the physical nature of memory, through system architecture and alignment ethics, to a reconception of AI safety centered on preserving the user's subjecthood rather than enforcing rules. Each principle is grounded in implementation experience rather than theoretical speculation.