Shopify's CEO Built a Search Engine That Revives 700 Claude Code Sessions in One Second

We live in an era where CEOs write code. Coinbase CEO Brian builds apps with Cursor. Shopify CEO Tobias programmed a search engine himself and pushed it to GitHub. That search engine is QMD, and when you connect it to Claude Code, every session gains persistent memory.

Artem Zhutov, a developer pursuing a Physics PhD, built this architecture in production. Over three weeks and 700 sessions, he discovered why the real bottleneck in AI coding isn’t model performance — it’s memory.

The Real Bottleneck Isn’t Performance — It’s Memory

Whether you use Claude Code or Codex, every new session starts from a blank slate. Artem ran 700 sessions over three weeks. Each time, he had to explain from scratch: “Here’s where we left off on this project.” Once context usage crosses 60%, the conversation gets compressed, and half the decisions made earlier vanish.

Model performance improves every month, but none of that matters if you can’t pick up where you left off yesterday.

• AI forgets earlier instructions as sessions grow longer
• Context compression structurally loses critical decisions
• Manually managing 700 sessions’ worth of context is impossible
• The problem isn’t the model — it’s the absence of a memory system

Everyone Takes Notes. Retrieving Them Accurately Is the Real Skill

Plenty of developers diligently pile notes into Obsidian. But accumulation alone is only half the equation. Notes become useful only when you can pull exactly what you need at the right moment. That’s the entire premise behind RAG — extracting just the relevant fragment from a mountain of stored data.

The default Claude Code approach uses a Haiku sub-agent that brute-force scans files. In practice, it takes three minutes to return 300 files, and almost none of them are useful.

QMD replaces this with three retrieval methods:

• BM25 search: Scores by word frequency and rarity, returning results in under one second
• Semantic search: Search “couldn’t sleep” and documents about “sleep improvement goals” surface
• Hybrid mode: Ranks relevance with 89% accuracy

Grep returns every sleep() function call when you search for “sleep.” QMD reads context instead. It handles abstract queries like “find ideas I never acted on,” and runs entirely locally so your data never leaves your machine.

One Line of /recall Brings Back Yesterday’s Entire Workflow

Layer a Claude Code skill called /recall on top of QMD and memory retrieval becomes automatic.

• /recall yesterday: Restores the previous day’s 39 sessions as a timeline
• /recall topic [keyword]: Collects every related file within one minute
• /recall graph: Visually explores a full week of sessions

When you close a session, it automatically parses the JSONL transcript and feeds it into the QMD index, keeping everything up to date.

The truly remarkable moment came when Artem searched “find ideas I never acted on.” Up came a note he’d written during a moment of near-surrender while working on his PhD thesis — a record he himself had forgotten. Manual search would never have surfaced it.

It’s No Longer Strange for a CEO to Write Code

Tobias requires every Shopify employee to treat AI as a baseline competency. He led by example, coding his own tool first. Brian is moving in the same direction.

Honestly, this isn’t about whether CEOs should be developers. They have far more pressing responsibilities. But the gap between people who can solve their own problems with AI and people who wait for someone else to build the tool is widening every month. Building things yourself is how you feel AI’s true destructive — and creative — power.

• Tobias built QMD and open-sourced it on GitHub
• Brian is building apps with Cursor
• Shopify reviews “Can AI replace this role?” before every hire
• QMD runs locally — no risk of company data leaving the building

The Only Thing That Persists Is Your Accumulated Context

Tools change every month. Nobody knows what comes after Claude Code. The only thing that persists is the context you’ve built up. People with a system for recording and retrieving will be productive from day one, no matter what tool arrives next.

QMD is the first building block of that system. Check it out on GitHub.