Self-Improving Agent

Problem: feedback rarely compounds

Agents often start every session from zero. Corrections, project preferences, and reusable workflows remain scattered in chat history. Over time this creates two risks: important rules are not retained consistently, or one-off context is mistaken for durable preference, making behavior unpredictable. self-improving focuses on turning corrections, preferences, and self-reflection into auditable local memory.

How it works: tiered memory and proactive maintenance

The skill organizes state under ~/self-improving/. Its core approach is not online learning, but maintaining readable and exportable Markdown rules. Typical steps include:

• Capture learning signals: when the user explicitly corrects the agent, states a stable preference, or when the agent identifies an improvement after significant work, record it in corrections.md or memory.md.
• Store by tier: frequent rules stay in the HOT layer of memory.md; project and domain patterns stay in projects/ and domains/ as WARM; low-frequency content moves to archive/ as COLD.
• Promote conservatively: a lesson is promoted only after repeated evidence, avoiding one-off instructions, hypothetical questions, or silence being treated as long-term rules.
• Cite sources: when using memory, name the source file; support querying, exporting, demotion, and loading the HOT layer first when context is tight.

Scope

It fits agent workflows that need stable memory for user preferences, project conventions, and correction lessons. It does not read calendars, email, or contacts, does not make network requests, and should not blindly delete memory during heartbeat cleanup. Sensitive data, health data, and third-party information should follow the boundary rules and not be written into local memory.