Self-Improving

A Claude Code/Codex skill for safe continuous improvement.

Attribution

This project is an independent Claude Code/Codex adaptation that combines ideas from:

• https://github.com/zhaono1/agent-playbook/tree/main/skills/self-improving-agent — structured multi-memory design, semantic pattern memory, confidence/application tracking.
• https://github.com/peterskoett/self-improving-agent — low-risk learning logs, hook reminders, markdown log formats, skill extraction scaffold.
• https://github.com/Toolsai/auto-skill — optional keyword-indexed knowledge base and cross-skill experience memory layer ideas.

The implementation here is rewritten with conservative safety defaults: raw local capture may be automatic, but promotion to durable guidance requires consent.

Storage Layout

By default, self-improvement data lives under the skill directory:

<skill-dir>/
└── data/
    ├── logs/
    │   ├── LEARNINGS.md
    │   ├── ERRORS.md
    │   └── FEATURE_REQUESTS.md
    ├── memory/
    │   └── semantic-patterns.json
    ├── knowledge-base/
    │   ├── _index.json
    │   ├── workflow.md
    │   ├── coding.md
    │   └── writing.md
    └── experience/
        ├── _index.json
        └── skill-self-improving.md

The initializer infers the skill directory from its own location: scripts/ parent directory is treated as the skill root. Optional project-local buffers can still be initialized under <project>/.self-improving/.

Quick Start

./scripts/init-learnings.sh
./scripts/init-learnings.sh --scope project --root /path/to/project
./scripts/init-learnings.sh --scope both --root /path/to/project

Windows/PowerShell:

pwsh -File scripts/init-learnings.ps1
pwsh -File scripts/init-learnings.ps1 -Scope Project -Root .
pwsh -File scripts/init-learnings.ps1 -Scope Both -Root .

The default scope is global and initializes <skill>/data/ without overwriting existing files. Project scope initializes <project>/.self-improving/ as an optional local buffer.

Optional Claude Code/Codex hook config examples live in settings/; replace <path-to-skill> with the absolute installed skill path before enabling them.

Consent Model

• Raw capture is automatic: failures, user corrections, non-obvious fixes, and feature requests can be recorded to the skill-local data/logs/ without asking each time.
• Pattern indexing is automatic: distilled summaries can update the skill-local data/memory/semantic-patterns.json without storing raw sensitive output.
• Promotion is consent-based: ask before writing to data/knowledge-base/, data/experience/, CLAUDE.md, AGENTS.md, Copilot instructions, new skills, commits, or remote pushes.

Design

• Hooks remind; they do not silently mutate durable guidance files.
• The default global store is skill-local data/, so it moves with the installed skill instead of living under the user home directory.
• data/logs/ is the raw event log, and data/memory/ stores redacted pattern summaries.
• data/knowledge-base/ and data/experience/ are promoted memory.
• Optional project-local .self-improving/ buffers are ignored by git by default.
• Knowledge and experience indexes are read only when relevant, not on every turn.