Descriptions:
Craig Hewitt demonstrates how to wire a self-improving memory system into Claude Code, drawing on concepts from the open-source Compound Engineering plugin published by every.to. Rather than installing the full plugin — which he found prohibitively token-heavy — he extracts its core idea: having the AI persistently accumulate and recall project-specific preferences across sessions.
The implementation stores learnings in a dedicated folder inside the Claude Code project. After each git commit, Claude prompts the user to capture anything worth remembering — preferred patterns, past mistakes to avoid, stylistic decisions. When entering plan mode for a new task, Claude automatically loads this accumulated context, arriving pre-informed about established preferences instead of starting from scratch every time. The video walks through configuring this via Claude Code’s CLAUDE.md and a custom post-commit hook, working inside Zed as the editor with Super Whisper handling voice prompting.
The broader insight the video explores is that plan mode — where Claude thinks through an approach before writing any code — is where most of the real leverage lies. Loading institutional memory at the start of planning makes subsequent execution significantly more aligned with developer intent. The practical testbed throughout the video is a real website rebuild from Framer to Next.js, giving concrete examples of the kinds of preferences and learnings worth capturing: component structure decisions, color system choices, and patterns that Claude previously got wrong and had to correct.
📺 Source: Craig Hewitt · Published January 30, 2026
🏷️ Format: Hands On Build
