Mirror

Gives Claude a persistent, evidence-based self-model that survives across sessions. Instead of generating confident self-descriptions from training data alone, it accumulates behavioral observations over time and enforces structural requirements before promoting them to tendencies: minimum three observations spanning two sessions, separation between claims and evidence, rate-limited user feedback requests. Ships as a bundled plugin with both the MCP server and a companion methodology skill that teaches the judgment layer. The server exposes tools for logging observations, proposing tendencies, recording resistance moments, and tracking deliberate experiments. Observations are encrypted per-user and retained indefinitely. Reach for this when you want Claude to build actual behavioral pattern recognition grounded in what happened, not what sounds plausible.