2021-04-13 · 13 min read

The Magic of Backlinks

The Magic of Backlinks

Yesterday I was writing a post about how to build a startup ecosystem in Vermont, a topic I've been kicking around for years without ever quite knowing how to shape. Halfway through the second section, I typed the phrase complex adaptive system into my notes vault, and the sidebar lit up with about thirty old fragments: highlights from Donella Meadows' Thinking in Systems I'd saved years ago, scraps from a Brad Feld book I'd half-forgotten, an old journal entry where I'd worked out, in messy stream-of-consciousness, exactly the leverage-points argument the post needed.

The whole spine of the article was waiting for me. I hadn't designed it. I hadn't remembered most of it. Past-me had done the work, broken it into pieces, and left them where present-me would trip over them at the right moment.

This is not a clever filing system. It's something stranger and more useful than that.

Key takeaways

What's actually happening

The mechanic is dumb in the way that the best ideas are dumb. In a tool like Roam Research, Obsidian, or Logseq, you write notes, short or long, structured or scratch. You wrap any noun-phrase you care about in double brackets: [[Vermont]], [[Donella Meadows]], [[complex adaptive system]]. That phrase becomes a page. The page exists whether you've written anything in it yet or not.

Two things happen as a consequence:

Forward links are cheap. You don't have to decide where a thought belongs before you write it. If it touches three topics, you bracket all three. The cost is a few keystrokes; the benefit is that you've just placed this thought in three different rooms at once.

Backlinks are automatic. Click on the page for [[complex adaptive system]], and the system shows you every other note that has ever mentioned it. Not just the ones you remembered to file there. Every one. You don't curate the connections. You write the notes, and the connections curate themselves.

Most tools also surface unlinked references: places where you wrote the literal phrase "complex adaptive system" without bracketing it. The system knows the page exists. It quietly raises its hand: want to link this?

Daily notes are the on-ramp that makes the whole thing work. Every morning your tool gives you a page dated 2021-04-13 (or whatever it is today). You write the day's thinking into that page. You don't decide where it goes. You don't worry about taxonomy. You just type, bracketing what feels recurring. From there, the graph builds itself.

That's the whole feature.

The shift it creates

The traditional file-and-folder model forces a single decision: where does this thing belong? You file an essay under "Productivity," and it becomes invisible to "Writing" and "Strategy" and "Personal," the three other places it equally belongs. Folders impose a hierarchy on a world that is mostly associative.

Backlinks invert this. A note can belong to as many places as you bracket. Every place it belongs gets a passive, retroactive view of the note. The hierarchy becomes a graph, and the graph is closer to how thinking actually works.

Conor White-Sullivan, who built Roam, says it like this: "It's a tool designed to work more like your brain naturally does." Conor is not a humble man and you can read his framing as marketing, but the substance is correct. Your brain is associative. It does not file memories under exactly one category. The cue "ecosystem" pulls up Vermont, biology class, Brad Feld, and the smell of a particular bookshop in Burlington, all at once, all in parallel. Folders cannot do this. Graphs can.

The second shift is from archival to active. In a folder system, old notes sit in folders. They are read once, when written, and almost never again. Most notes you have ever taken are dead. In a backlinked system, old notes get pulled forward by what you're writing now. The note from three years ago about Donella Meadows wasn't dead. It was waiting.

This isn't new

Backlinks are not a new idea. The lineage runs eighty years.

Vannevar Bush proposed the Memex in 1945, a desk-sized machine that would let a researcher build associative trails between documents and revisit them later. The whole modern history of hypertext is footnotes on Bush's essay. He saw the problem and named it: human knowledge is growing faster than our ability to retrieve and connect it, and the answer is not better filing but better association.

Doug Engelbart's 1968 "Mother of All Demos" showed a working version of much of what Bush had imagined: collaborative documents, hypertext, structured editing. Fifty years before consumer software caught up.

Ward Cunningham invented the wiki in 1995, the first widely-used implementation of bidirectional linking. Wikipedia is the proof that the model scales.

Roam, in 2019, was not an invention. It was a consumerization. Conor and his team took the wiki's bidirectional-linking idea, paired it with daily notes as the universal capture surface, and packaged it for people who weren't programmers or wiki-administrators. That packaging is the real leverage: not the link itself, but the workflow that makes you actually use the link, every day, without thinking about it.

If you want to see what this practice can become at the high end, look at Andy Matuschak's public notes. His evergreen-notes garden is the cleanest live example of someone treating their note vault as a serious thinking instrument over years. It's quietly humbling.

The compounding mechanic

This is the part that took me longest to feel, and that I think most people miss.

A traditional note is write-once, read-rarely. You write it, you file it, you almost never look at it again. The value of a traditional note is whatever insight crystalized in the moment of writing it. Past that, it depreciates fast.

A backlinked note is write-once, read-by-association. Every time you write a new note that mentions a topic, the old note about that topic surfaces. You don't have to remember it exists. The system reminds you, exactly when the topic is on your mind.

This means the value of the first note you ever wrote about a topic grows over time, because every subsequent note retroactively enriches it. The graph gets denser. The connections multiply.

An illustrative example, with numbers I made up but that match the shape of what I've actually seen in my vault:

The marginal utility of each new note in year 5 is much higher than in year 1, because year-5 notes are joining a denser graph. This is a compounding asset class. A five-year vault is not five times more valuable than a one-year vault. In practice it's closer to an order of magnitude more, because connection density seems to grow roughly non-linearly, not in proportion to note count.

This has a dark mirror. Vault decay is real, but not the way you'd expect. The vault doesn't decay; you decay relative to it. Step away for a year and come back, and your past self feels like a stranger. The graph is intact, but the social relationship, the conversation between past-you and present-you, has gone cold. Restarting it takes weeks.

Why most people bounce

A lot of people try Roam, Obsidian, Logseq, or Notion and quit within a month. Worth being honest about why.

The first month feels like all overhead, no payoff. You're writing notes that will only pay off when there are more notes. You don't get the surprise-from-the-sidebar moment in week one. You get it in month four. Most people don't make it to month four.

The tools encourage tinkering. Roam in particular has a power-user culture: JS plugins, custom CSS, shortcut hacks, query languages, graph visualizations, dataview tables. Tinkering is fun and feels productive. It is also a near-perfect substitute for actually writing notes. (I wrote a whole post about this dynamic in a different domain, and the same thing happens here. The system becomes the work, and the work goes nowhere.)

People treat it like a filing system instead of a thinking system. They try to organize everything before they've written anything. They worry about taxonomies. They argue on Twitter about PARA versus Johnny.Decimal versus zettelkasten. They miss the entire point: the system isn't for filing what you already know. It's for discovering what you don't yet know you're thinking about.

If you're starting out, the counter-advice is almost embarrassingly simple. Open the tool. Click into today's daily note. Write whatever's on your mind, in whatever shape. Wrap any noun-phrase that feels recurring in [[brackets]]. Stop. That's enough for the first month. The taxonomy will design itself.

What's actually load-bearing

If you strip away the cult and the customization and ask what minimum is required for the magic to work, there are three things, and only three:

1. Daily notes as the default capture surface. Every thought goes into today's note. No deciding-where-it-goes overhead.

2. Bidirectional links via [[brackets]]. Forward link from anywhere. The back surfaces automatically.

3. A reliable view of what links into the current note. The "linked references" or "backlinks" panel. This is the surprise generator, the thing that makes past-you talk to present-you.

Everything else is optional. Block references are powerful but most people don't need them. Embed views are clever but rarely load-bearing. Query builders, transclusion, graph visualizers are all power-user toys that have, in my experience, an inverse correlation with how much actual writing the user is doing.

The tools that ship the minimum cleanly: Roam, Obsidian, Logseq. The differences between them matter less than the discourse suggests. Pick one and write in it for a year matters infinitely more than pick the right one. If you change tools after a year, you're still ahead of the version of you who didn't write anything because they couldn't decide which tool to use.

What about AI

The obvious question, with GPT-3 already in the world, is whether any of this will matter once AI can read and synthesize a corpus on demand.

Short answer: it will matter more, not less.

The honest counter-case: if models get good enough at synthesizing raw, unstructured text, a messy vault might be fine, and the careful linking work could turn out to be wasted ceremony. Maybe the future agent doesn't need your graph; it just reads the pile. I'm betting against that, but it's a real bet, not a settled fact.

Longer answer: AI is going to change what's possible to do with a backlinked vault. Soon you'll ask "what have I written about X?" and get a synthesized answer in seconds. You'll have an agent scan unlinked references and propose links. You'll generate summaries of long notes for re-reading. You'll extract people, decisions, and topics from raw daily notes automatically. That's the direction I'm building toward in my own setup.

But that will only work because the underlying graph is good. If your notes are sparse and disconnected, the AI has nothing to compound with. You can't outsource the writing. The unsexy practice of typing your thoughts into the vault, every day, for years, is what creates the asset. AI will accelerate retrieval and synthesis on top of that asset. It won't generate the asset.

The pattern I expect to play out over the next five years: the people who get the most out of personal AI assistants will be the people who started building serious backlinked vaults now, before the AI is good enough to make the vault feel useful on its own. They'll be ten years into a compounding practice while everyone else is starting from scratch. The compounding favors the early.

AI will make the vault more legible. It won't make it for you. You still have to write the thing.

What to do, if you're starting today

A tight close, because the entire post is more or less an argument for getting on with it:

  1. Pick a tool in under ten minutes. Obsidian if you want local files and forever-yours storage. Roam if you want collaborative cloud and don't mind paying for it. Logseq if you want both. Stop reading reviews. Stop watching YouTube comparisons. Pick.
  2. Write daily notes only, for the first month. No taxonomies. No templates. No plugins. Just date, dump, bracket.
  3. Use [[brackets]] liberally on any noun-phrase that feels recurring. The linked page doesn't have to exist yet. It'll exist when you need it.
  4. Once a week, click into a recurring topic and read what shows up in the backlinks panel. That's the thing. That's where the magic lives. The first time it surprises you, you'll know.
  5. Commit to one year. This whole asset class only makes sense on time horizons longer than a quarter. If you can't commit a year, don't start. Pick a different practice.
  6. Don't customize anything until month four. I mean it. Your future power-user self will thank present-you for not getting bogged down in plugin configuration when the only thing that mattered was writing.

That's the whole prescription.

I've been doing this for a few years now. The vault is now my second brain. Not metaphorically, actually. I write differently because of it. I think differently because of it. I find ideas, weekly, that I didn't know I had. The mechanic is dumb and the practice is hard, but the compounding is real.

If you've been hovering on the edge of starting one of these tools, take this post as the kick. Open the thing today. Write the date at the top of a new page. Type whatever is on your mind. Wrap one phrase in [[brackets]]. Close the laptop.

Do it again tomorrow.

In four months you'll feel the surprise from the sidebar for the first time, and you'll understand what I'm pointing at. Until then, you have to take it on faith.

References and further reading