Skip to content
Book Open access

MemCanvas: A Structured Memory Canvas for Evolving Intent Alignment in LLM Dialogues

Nov 2026 · ACM Symposium on User Interface Software and Technology · 0 citations · 74 references
Computer Science

Abstract

Aligning user intent in complex, evolving dialogues with Large Language Models (LLMs) is a significant challenge. We attribute this to a cognitive divergence between LLM’s linear reasoning based on context and statistical semantic representations, and the exploratory, non-linear, and personalized nature of human intent. These gaps not only cause progressive misalignment but also leave users without a clear understanding of model’s perception or means to control its trajectory. To bridge this gap, we introduce MemCanvas, an interactive system that organizes dialogue history into shared, editable memory structures to complement outcome-based approaches. By scaffolding intent via three interaction primitives—Trees (for thematic directions), Relations (for logical connections), and Tags (for personalized semantics)—MemCanvas enables users to externalize evolving intent and proactively steer the LLM’s reasoning. Our technical evaluation and user study (N=28) demonstrate that MemCanvas effectively aligns user intent and enhances users’ sense of control, transparency and collaboration. Ultimately, this work illustrates how intent scaffolding can serve as a practical supplement to standard linear interfaces, fostering more process-aligned human-AI collaboration.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.