Contextual Memory Virtualisation: DAG-Based State Management and Structurally Lossless Trimming for LLM Agents

Researchers introduce Contextual Memory Virtualisation (CMV), a system that preserves LLM understanding across extended sessions by treating context as version-controlled state using DAG-based management. The system includes a trimming algorithm that reduces token counts by 20-86% while preserving all user interactions, demonstrating particular efficiency in tool-use sessions.

• →CMV treats LLM accumulated understanding as version-controlled state that can be preserved across context limit resets.
• →The system uses a Directed Acyclic Graph (DAG) to model session history with snapshot, branch, and trim capabilities.
• →A three-pass trimming algorithm reduces token counts by mean 20% and up to 86% while preserving all user messages verbatim.
• →Testing across 76 real-world coding sessions showed strongest gains in mixed tool-use sessions averaging 39% reduction.
• →The system reaches break-even within 10 turns and remains economically viable under prompt caching.