CogCanvas: Verbatim-Grounded Artifact Extraction for Long LLM Conversations

CogCanvas is a training-free framework that improves long conversation memory in large language models. Instead of compressing chat history, it extracts and retrieves literal artifacts (decisions, facts, insights) using a temporal-aware graph, mimicking how teams use whiteboards. On the LoCoMo benchmark, CogCanvas achieves 32.4% accuracy, outperforming RAG by 7.8 percentage points, with significant gains in complex reasoning (+20.6pp on temporal reasoning).