Token Warden optimizes Claude Code memory

The Token Warden plugin automates context management for AI agents. It treats agent memory as an engineering problem: each rule must prove its effectiveness on benchmarks to stay in the system.

The system works in four stages. It collects session data, analyzes it, and proposes rules. These rules are then tested on a reference repository, and only those that save tokens with a margin are selected. Rules that no longer pay off are automatically removed.

The plugin supports integration with Claude Code v2.1 and above. It allows users to track rule effectiveness with the /warden-status command and conduct A/B testing of prompts or models. The entire learning process happens without blocking the user’s main work.