[FEATURE] Self-improving learning loop — autonomous skill creation from experience (like Hermes Agent)

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Claude Code has no memory of how it solved problems in past sessions. Every time I start a new session, it has to figure out the same things from scratch — even for tasks I've run dozens of times. There's no way for it to learn from successful completions, remember my workflow preferences, or reuse approaches that worked well before. This makes it slower and less personalized over time, when it should get smarter.

Proposed Solution

I'd like Claude Code to implement a self-improving learning loop similar to what Hermes Agent (by Nous Research) does:

1. After completing a complex task (e.g. 5+ tool calls), Claude Code should automatically extract the approach as a reusable "skill" — a structured document with the procedure, pitfalls, and verification steps.
2. On similar future tasks, it loads that skill instead of reasoning from scratch.
3. Skills should self-improve when the agent discovers a better approach during use.
4. A persistent user model should build over sessions — remembering preferences, coding style, project conventions — without needing re-instruction each time.

This does NOT require changing the underlying model. It's purely an agent-layer memory and skill persistence system.

Alternative Solutions

I've tried writing CLAUDE.md files manually to persist instructions, but this requires me to do all the work. I also tried custom slash commands, but they don't adapt or improve automatically. Hermes Agent (github.com/NousResearch/hermes-agent) solves this at the framework level — Claude Code could adopt a similar approach natively.

Priority

Critical - Blocking my work

Feature Category

Configuration and settings

Use Case Example

Example scenario:

1. I frequently debug a Python FastAPI app with a specific project structure
2. Claude Code currently re-discovers the same patterns every session (where configs are, how I run tests, how I handle migrations)
3. With a learning loop, after the first few sessions it would create a skill: "FastAPI project workflow for this user"
4. Future sessions would load that skill instantly, saving 5-10 minutes of re-explanation per session and producing better first attempts

Additional Context

Reference: Hermes Agent by Nous Research — https://github.com/NousResearch/hermes-agent
It uses a 4-layer memory system (episodic + procedural + user model + always-on context) and autonomous skill creation after complex tasks. Skills are stored as markdown files and self-improve during use.

This feature would make Claude Code the only major AI coding CLI with a genuine built-in learning loop — a significant competitive advantage over Cursor, Copilot, and others.