![[2602.22402] Contextual Memory Virtualisation: DAG-Based State Management and Structurally Lossless Trimming for LLM Agents](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.22402] Contextual Memory Virtualisation: DAG-Based State Management and Structurally Lossless Trimming for LLM Agents
Summary
The paper presents Contextual Memory Virtualisation (CMV), a novel system for managing state in large language models (LLMs) using a Directed Acyclic Graph (DAG) structure to enhance context reuse and reduce token counts through a lossless trimming algorithm.
Why It Matters
As LLMs are increasingly used for complex reasoning tasks, managing their state effectively becomes critical. CMV addresses the limitations of current context management by preserving essential information while optimizing resource usage, making it relevant for developers and researchers in AI and software engineering.
Key Takeaways
- CMV employs a DAG structure for effective state management in LLMs.
- The three-pass trimming algorithm achieves a mean token reduction of 20%, with up to 86% in specific cases.
- The system maintains verbatim user and assistant interactions, ensuring no loss of critical information.
- A case study shows significant efficiency gains, especially in sessions involving mixed tool use.
- CMV's approach can enhance the economic viability of prompt caching in LLM applications.
Computer Science > Software Engineering arXiv:2602.22402 (cs) [Submitted on 25 Feb 2026] Title:Contextual Memory Virtualisation: DAG-Based State Management and Structurally Lossless Trimming for LLM Agents Authors:Cosmo Santoni View a PDF of the paper titled Contextual Memory Virtualisation: DAG-Based State Management and Structurally Lossless Trimming for LLM Agents, by Cosmo Santoni View PDF HTML (experimental) Abstract: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 remai...
