PIA/DMA: Persistent Identity and Dual-Mode Autonomous Scheduling for Long-Term AI Agents

Abstract

Large language model agents today face two fundamental limitations: they forget who they are between sessions, and they do nothing between responses. We present the Persistent Identity Architecture (PIA) and the Default Mode Analog (DMA) — two complementary subsystems that together address both limitations within a single agent runtime. PIA provides a five-layer memory architecture (episodic anchors, semantic fact store, identity profile, coherence engine, provider abstraction) that maintains coherent agent identity across session boundaries, model provider changes, and computational substrate transitions. DMA provides a scheduled dual-pulse cognitive rhythm — directed work sessions and undirected curiosity sessions — that enables autonomous productivity and emergent knowledge recombination during user-offline periods. Deployed together in production over 20+ consecutive days, the combined system produced over 250KB of autonomous research output, maintained identity coherence across multiple model provider switches, and demonstrated unprompted cross-domain associative connections from internal representation alone. We present the architecture, operational data, and design principles for building agents that persist.

Keywords