Field notes · 18 chapters
What it costs to build a workspace
that remembers.
What broke, what the fix actually was, and the parts that don't flatter us. Written while it was still fresh enough to be honest about.
Origin
Why this exists at all.
Decisions
The forks, and what we picked.
- 01The case for not specialisingEvery good productivity app picks one job and does it well. We picked all of them, and the seams are the whole point.
- 04The assistant is a user, not a featureThe chat box is the least interesting way to put AI in a product. We gave the assistant the same door everyone else uses.
- 05A shared brain needs walls before it needs anything elseWe built permissions in phase two, long before anyone was sharing anything. A family workspace without them is a different, worse product.
- 06The runtime that lives somewhere elseThe assistant runs in a separate Python service on its own machine. That boundary has cost us real effort, and it is the reason the rest of the product still behaves like software.
- 07What is worth remembering, and who decidesAn assistant that remembers everything is surveillance. One that remembers nothing is a stranger every morning. The interesting work is the line between them, and who gets to move it.
- 08Asking for lessGoogle rejected our OAuth verification for asking too broadly. They were right, and the narrower scopes made the integration better.
Mechanism
How the thing actually works.
- 02Airplane mode is not an error stateWe put a real database in your browser in week four, before there was much product to put in it. Then spent months finding out we'd been testing offline wrong.
- 03Two planesYour data lives on your device. Your billing tier does not. Living with that split costs six files per table, and we have the scar to prove it.
- 09Teaching the Mind to stop connecting breakfast to breakfastA screenshot of our own noisy review queue, a chip company's knowledge base, and the week we rebuilt how Margin connects your ideas.
- 11The day the graph learned who your people arePerson-to-person edges, duplicate identities, and a text box that lets you correct the machine — three weeks of making the Mind connect like a brain instead of a database.
- 13The note that was 98% invisibleOur founder had a feeling the AI's understanding was 'half-baked'. He was right: a 238,000-character note had been summarized from its first 4,000. Here's the fix — and why gut feelings deserve queries.
- 16The assistant that talked itself in circlesAsked to move a workout plan by one week, our assistant said 'let me do it now' two hundred times and never did it. The autopsy vindicated the model we almost blamed.
Trust
What broke, and what it cost.
- 10The vault that leaked through its own wallsWe built an encrypted vault the AI cannot read. Then we found 110 pieces of sealed content woven openly through the product — and learned what containment really means.
- 12The memories that never happenedOur AI's remember function had silently failed on every single call, forever — and every safety net we built is exactly why nobody noticed.
- 14The day we refused to defer anythingOne founder's feedback loop, eleven shipped systems, a brutal self-audit, and the two words that shaped it all: nothing deferred. A chronicle of the longest day in Margin's engineering life.
- 15A brain that knows what changedNotes don't grieve. Task lists don't notice a wedding. The hardest problem in personal AI isn't remembering — it's letting knowledge live, age, and change tense the way a life does.
- 17The face in the marginMargin's intelligence needed a body — something to blink, squint, and worry so a spinner never has to. On drawing a character in three strokes and finding its soul in a horizontal line.