I built a Claude Code plugin that only keeps an "agent memory" rule if it can prove it saves tokens

Short summary

Developer open-sourced token-warden, a Claude Code plugin that validates agent memory rules through token benchmarking. Rules must save 2x their overhead or auto-evict. Real case: a grep-first rule cut wasteful agent sessions 16% (67k to 56k tokens, ~3¢/run on Sonnet), but vanished in optimized workflows—measurement over vibes.

• •Open-source tool that benchmarks Claude Code agent memory rules for real token efficiency
• •Rules auto-evict unless they save at least 2x the tokens they cost to maintain
• •Measured example: grep-first rule saved 16% on wasteful agents, disappeared on optimized ones

Generated with AI, which can make mistakes.