Self-Improving Agent

Problem

AI agents often repeat the same failures across multi-turn tasks, projects, and parallel sessions: user preferences get ignored again, error context disappears after the current conversation, reusable patterns are not preserved across projects, and accumulated experience rarely becomes executable skill. self-improving addresses this gap by adding a lightweight memory and review loop to the agent, instead of relying on the model to reconstruct context from scratch every time.

How It Works

The skill works through capture, reflection, sharing, and extraction. At session start, it performs a light sweep: removes files in ~/self-improving/shared/ older than 7 days, keeps the shared directory under 20 files, and compacts memory.md when it exceeds 100 lines. During work, it detects user corrections, runtime errors, feature gaps, knowledge gaps, and better approaches, then stores reusable information in project-local or global files. For cross-session exchange, it uses a flat shared/ directory as a lightweight board: each file contains one short insight, and other sessions can read it at task start or context switches, apply it, and delete it after use. Memory is organized into HOT, WARM, and COLD tiers, with frequent rules kept in memory.md, project and domain patterns stored in more specific files, and stale content demoted to archives. Once a pattern is repeatedly validated, it can be extracted into a SKILL.md.

Boundaries

It fits long-running collaboration, multi-project reuse, and cross-session consistency in engineering workflows. Be careful to record only explicit corrections or verified issues, avoid treating silence as rules, and never store secrets, tokens, or sensitive personal data. Do not load all memory at once; prefer the HOT tier and the smallest matching files. Shared notes should capture reusable rules, not debugging details from a single session.