Two-Room Memory Architecture: Efficient Context Management for LLM Memory Systems via Triviality Gating

Current approaches to LLM memory management lack principled mechanisms for distinguishing meaningful user information from conversational noise. We propose a two-room architecture with unidirectional information flow, where all input enters an active buffer (Room 1) and is evaluated by a triviality gate before persisting to long-term storage (Room 2). The key insight is an inversion: rather than asking "is this information important enough to store?" the system asks "is this information trivially dismissible?" This reframing exploits the observation that triviality forms a tighter semantic cluster than meaningfulness. We validate this approach using a classifier trained on 113 labeled examples, achieving 100% accuracy on held-out novel examples. We further propose organizing persistent memory by relational posture (what the information demands from the system) rather than data category, and introduce a two-tier storage model based on mutability likelihood. This architecture offers significant efficiency gains by reducing storage requirements, improving retrieval relevance, and providing a principled framework for LLM memory management.