Working memory for Claude Code – persistent context and multi-instance coord (github.com)

Claude Cognitive has introduced a groundbreaking enhancement to Claude Code, equipping it with working memory capabilities that allow for persistent context and effective multi-instance coordination. Traditionally, Claude Code functions statelessly, causing it to start afresh with each session, leading to repetitive tasks and inefficient integrations, especially in large codebases. The new system addresses these issues through an attention-based file injection mechanism categorized into HOT, WARM, and COLD states, dynamically managing file context based on user interactions. Additionally, it supports multi-instance state sharing, enabling automatic or manual coordination among instances, which reduces token usage significantly—up to 79% savings in cold start scenarios. This advancement holds considerable significance for the AI/ML community, particularly for developers handling substantial and complex codebases. By ensuring new instances can maintain productive continuity without hallucinations or repeated tasks, Claude Cognitive not only simplifies the development workflow but also enhances collaboration across multiple teams. With successful validation on production codebases exceeding 1 million lines, this tool promises to redefine how developers interact with AI models, making long-term memory retention a practical reality in coding environments.