Search papers, labs, and topics across Lattice.
The paper introduces Contextual Memory Virtualisation (CMV), a system for managing LLM context as version-controlled state to mitigate information loss during context window compaction. CMV represents session history as a Directed Acyclic Graph (DAG) with snapshot, branch, and trim primitives, enabling context reuse across parallel sessions. The authors propose a three-pass structurally lossless trimming algorithm that reduces token counts by stripping mechanical bloat while preserving user messages and assistant responses, achieving up to 86% reduction in token usage.
LLMs can retain more context and history without blowing your token budget using a DAG-based memory system and lossless trimming that cuts context length by up to 86%.
As large language models engage in extended reasoning tasks, they accumulate significant state -- architectural mappings, trade-off decisions, codebase conventions -- within the context window. This understanding is lost when sessions reach context limits and undergo lossy compaction. We propose Contextual Memory Virtualisation (CMV), a system that treats accumulated LLM understanding as version-controlled state. Borrowing from operating system virtual memory, CMV models session history as a Directed Acyclic Graph (DAG) with formally defined snapshot, branch, and trim primitives that enable context reuse across independent parallel sessions. We introduce a three-pass structurally lossless trimming algorithm that preserves every user message and assistant response verbatim while reducing token counts by a mean of 20% and up to 86% for sessions with significant overhead by stripping mechanical bloat such as raw tool outputs, base64 images, and metadata. A single-user case-study evaluation across 76 real-world coding sessions demonstrates that trimming remains economically viable under prompt caching, with the strongest gains in mixed tool-use sessions, which average 39% reduction and reach break-even within 10 turns. A reference implementation is available at https://github.com/CosmoNaught/claude-code-cmv.
