What is juice?

Juice is a small server that gives AI coding agents a persistent memory of things they should not do again. Most agent memory systems load everything into context at once, which gets noisy. Juice takes a different approach: it stores only negative guidance, meaning corrections and avoidance rules, and surfaces them selectively when they are relevant to the current task. The typical workflow is: a user catches the agent making a recurring mistake, the agent proposes an avoidance constraint in plain language, the user approves or edits it, and Juice saves it. Future sessions retrieve only the constraints that match what the agent is currently working on, keeping prompts focused. Each saved constraint includes the avoidance statement itself, a scope (global, project, repo, or agent), a category such as general, design, or writing, trigger words that control when it surfaces, and confidence and strength signals for ranking. The system deliberately rejects positive preferences and instructions that say always do this, keeping the store focused on what to avoid. Juice runs as either a local command-line server or an HTTP server for network access. Data is stored in a local SQLite file by default. It exposes a small set of tools through the MCP protocol: getting a manifest, preparing relevant constraints for a task, suggesting new constraints, saving confirmed ones, updating or retiring existing ones, and listing all stored constraints. A mobile-friendly web UI is also available for browsing and managing constraints. The server works with OpenCode, Claude Code, Cursor, Codex, and other MCP-capable agent clients. A bundled skill file can be installed into those clients with a single command.