Self-Improving + Proactive Agent

To make an agent learn from corrections, you need a persistent memory system that logs user feedback and promotes recurring patterns into confirmed rules. This skill maintains a local memory of preferences and lessons across global, domain, and project namespaces to improve execution quality over time.

Local agent memory works by storing state in a local directory using markdown files for hot, warm, and cold tiers. It preloads relevant rules before non-trivial work and uses a heartbeat-driven maintenance workflow to safeguard learning and compound execution quality across repetitive tasks.

Yes, you can manage agent memory separately using namespaced structures. This skill organizes memory into global, domain, and project namespaces, allowing the agent to apply specific rules and lessons contextually during non-trivial tasks without cross-contaminating preferences.

The best way to stop repeating mistakes is to log them explicitly in a corrections file. This skill tracks these corrections, identifies recurring patterns, and promotes them to confirmed preferences in its memory system, actively preventing the same errors in future workflows.

No, you do not need network access to maintain persistent memory. This skill operates entirely offline, storing state and managing heartbeat-driven maintenance locally on your machine to ensure data privacy and uninterrupted execution.

To archive decayed patterns, the memory system automatically moves outdated rules into a cold archive tier. This skill manages the lifecycle of learned preferences, moving confirmed rules from hot memory to cold archives when they are no longer actively needed.