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

Le résumé fourni par la source

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.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.