Self-Improving + Proactive Agent

Problem

Long agent conversations have a memory gap: the agent repeats mistakes, loses stated preferences, or overfits one-off instructions. Manual prompt maintenance does not scale. This skill turns explicit user corrections, confirmed preferences, and self-reflection into local memory files, so useful patterns compound over time instead of disappearing when context resets.

How It Works

• Learning signals: captures explicit corrections (“No, that's not right...”), direct preferences (“Always do X for me”), and repeated useful patterns; ignores one-time instructions, context-specific details, and hypotheticals.
• Tiered storage: memory.md is HOT and always loaded, capped around 100 lines; projects/ and domains/ are WARM and loaded on context match; archive/ is COLD and loaded only on explicit query.
• Promotion and decay: a pattern used 3 times in 7 days can move to HOT; unused for 30 days drops to WARM; unused for 90 days moves to COLD, with deletion only after confirmation.
• Conflict resolution: more specific namespaces win, project > domain > global; same-level conflicts use recency; ambiguous cases ask the user.
• Transparency: memory-driven behavior should cite the source, for example “Using X (from projects/foo.md:12)”.

Boundaries

It reads only files under ~/self-improving/, makes no network requests, and does not access calendar, email, contacts, or modify its own SKILL.md. It should not store credentials, health data, or third-party information. When context is tight, it should load HOT first and explicitly state what remains unloaded rather than failing silently.