What problem does it solve? AI agents repeat the same mistakes across sessions because they lack persistent memory of user corrections and preferences. This Skill gives the agent a structured, self-organizing memory so it learns from corrections, self-reflection, and repeated patterns instead of starting from zero every time. ## Core Features & Use Cases - Tiered Memory Architecture: Stores learnings in hot (memory.md), warm (projects/, domains/), and cold (archive/) layers under ~/self-improving/, with automatic promotion and demotion based on usage frequency. - Correction & Preference Learning: Detects explicit correction signals ("no, do X instead", "I told you before..."), logs them to corrections.md, and promotes patterns to confirmed rules after three repetitions with user confirmation. - Self-Reflection Workflow: After significant work, the agent evaluates results against intent and logs lessons in a structured format for future retrieval. - Safety Boundaries: Never stores credentials, financial data, medical info, or third-party information, and supports full memory export and wipe on request. - Use Case: A user repeatedly corrects the agent to use SQLite instead of Postgres for MVP projects; after three corrections the agent asks for confirmation, then permanently applies the preference and cites its source when used. ## Quick Start Ask the agent to set up the self-improving memory system and start learning from your corrections and preferences.
