Self-Reflection Memory

Problem

Agents in long conversations often repeat the same mistakes: users repeat preferences, fixed bugs resurface, and project conventions are not retained. Dumping all lessons into long context can exceed limits and can turn one-off instructions into durable rules. This skill turns corrections, preferences, and reusable patterns into auditable local memory instead of relying on implicit recall.

How It Works

It stores state under ~/self-improving/ with tiers: memory.md as HOT for confirmed high-frequency rules, projects/ and domains/ as WARM for scoped conventions, and archive/ as COLD for decayed material. It triggers logging when users explicitly correct behavior, state preferences, or after significant work. Candidate patterns usually need about three repetitions before becoming stable rules, avoiding inference from silence or one-off observations.

When users ask what the agent knows, show patterns, or request memory stats, it retrieves the relevant tier and reports counts. If files approach limits, it merges similar rules, archives stale patterns, and compresses verbose entries, while preserving confirmed preferences. Memory usage should cite sources for traceability. It only reads and writes its local directory, does not access calendar, email, or contacts, and makes no network requests. If context is low, it should degrade to HOT memory and disclose what is not loaded.