The browser problem nobody is solving

The Browser Company raised $128 million building Arc and Dia. Zen forked Firefox. They all missed the same thing.

Last Sunday I closed my browser to free up RAM. I had been deep in research for three days, chasing a question about attention mechanisms across twenty or so papers, GitHub threads, and scattered Stack Overflow answers.

Tuesday morning I reopened it. Arc restored my tabs. Technically. A third of them were there. Two thirds showed that sad little "restore" dropdown I've clicked a thousand times. I clicked it. Some restored. Some didn't. The thread I was pulling on, the reason I had opened the first tab, was gone. Not the URL, the URL was probably in my history somewhere. The thread. Why those tabs belonged together. Where I had gotten to in my thinking. What I had decided to come back to. What I had decided to skip.

I did the thing every knowledge worker does. I told myself I would remember, opened a new tab, and started again from Google.

This happens to me three times a week.

It's not a tab management problem. It's a memory problem. And a decade after the browser wars "ended", with more money pouring into browser startups than any time since Netscape, nobody is solving it.

What Arc actually built

I like Arc. I was a paying user for most of 2024. The sidebar is the best in the category. Spaces are real. The command bar is great. The Windows launch was messy but the macOS app is polished.

Here is what Arc is: a much better window manager for a browser. That's the innovation.

The Arc team looked at the Chrome tab bar, saw 30 years of "let's just keep adding pixels to the top", and said no. They built a vertical sidebar, they called tabs Pins and Favorites, they added Spaces for project separation. Every one of these moves is right. Arc makes it easier to keep what's on screen organized.

But close Arc. Reopen it. Your Space comes back, mostly. The tabs you had open, mostly. Not your scroll positions. Not why those tabs belonged together. Not the page you read twice and was going to bookmark. Not the other three tabs you had in a different Space that are relevant because they informed the decision you made in this one.

Arc organizes what's on your screen right now. It does almost nothing for what you did yesterday.

What Dia is betting on

Dia, from the same team, went the other direction. They put an AI agent in the URL bar. Smart bet. If you can't remember what you read, maybe you can ask. "Summarize this page." "Write me a response." "Pull the key numbers from these three tabs into a table."

It's useful. For about 30 seconds.

Then you ask "what was I reading about transformers last week" and Dia can't answer. Because the index doesn't exist. Dia can reason over what you're looking at right now, or what you've opened recently. It can't reach into your history, pull content, and find the passage. Nobody can. Browsers don't index what was on the page, only the URL and title.

Dia is a better tab. It's not a better browser.

What Zen promised

Zen is the Firefox fork. Open source, community driven, beautifully designed. I love it on principle. The calm aesthetic is a genuine achievement. They ship fast.

But open Zen, use it for a week, and notice what you're doing. You're managing tabs. You're closing windows because you have too many open. You're losing things in workspaces you forgot you had. The UI is prettier. The underlying model is still the 1994 one: tabs and windows, windows and tabs.

When you drag a tab to a workspace in Zen, the workspace is now "the set of tabs you happen to have there." Close Zen. Reopen it. It mostly works. Not entirely. I've lost sessions twice in a month. Not Zen's fault exactly. It's Firefox's session manager, which has always been a best-effort thing.

The pattern across Arc, Dia, Zen, Brave, Vivaldi, SigmaOS, Sidekick: surface innovation on top of a browser model that was designed before the web had Google, before knowledge work happened in twelve tabs at once, before anyone had ever heard of "context switching" as a cognitive cost.

What you actually do all day

Think about what happens when you research something seriously. Not browsing. Research. Something like "should we use server components for this feature" or "is there a better way to do batch embeddings" or "which visa do I need for this trip".

You open one tab. It leads to three. Two of those lead to a page each. One of those lands you in a thread that's five layers deep. You follow it. You come back. You go sideways. You read something that contradicts what you just read. You open a second perspective. You end up with fifteen tabs, five of which are useless (you know this, you'll close them), eight of which are in your main line of reasoning, and two of which are off to the side because they raised a question you want to address later.

Your context at this moment isn't those fifteen tabs. It's:

• Which of the fifteen are the spine of your argument
• Which three you read three times because they mattered
• The sentence in tab seven that reframed everything
• The mental checkpoint you set when you said "OK, I've decided X, now I just need to verify it"
• The parallel thread you started yesterday that's relevant here but you haven't realized yet

Close the browser. Every browser that exists today keeps the URLs. They keep maybe the scroll positions. A good one keeps the tab groups. None of them keep the reasoning shape. None of them index the content. None of them can tell you "you read that phrase last week, here's the page, here's the context you were in when you read it".

You are expected to manage this complexity with tabs. Tabs are linear. Your thinking isn't.

Why nobody builds this

Three structural reasons, not because nobody noticed.

The browser is infrastructure. Almost all browser R&D goes into the engine. V8 gets faster. CSS gets new features. Security gets tighter. The UI shell, the part a user actually experiences, gets a coat of paint every five years. The real experts in the space are working on rendering pipelines, not on thinking tools. The founders who could build a better shell go build Notion or Linear instead, because browsers are perceived as "done."

Arc showed the thesis stops working. Arc raised a huge round on "browser, reimagined." Two years later, the team pivoted to Dia. That's the real signal. Even the best-funded new browser team in twenty years decided that a better UI alone doesn't get you to escape velocity against Chrome. They went horizontal into AI. They didn't go deeper into context because context is expensive and the product model for it is unclear.

Local-first is economically hostile. To preserve context you have to store it. You have to index page content. You have to run full-text search on tens of thousands of pages locally. You have to keep graphs of navigation. You have to trust the user's machine. This is the opposite of what VC-backed browsers want. Cloud-first means sync revenue, data for training, lock-in. Local-first means none of that. The math doesn't work for a Series A pitch.

So the space where users most need help got abandoned at the exact layer where it matters.

What solving it actually requires

I'll describe what the fix looks like, in boring technical terms, because the interesting part isn't the vision, it's that none of this requires new research.

You need three things.

A navigation graph, not a history list. Every page you visit is a node. Every click is an edge. Branches when you open a link in a new tab. Forks when you go back and go a different direction. You can see the shape of your exploration. You can name a node ("before legal review"). You can jump to any node and reopen it with context. This is dagre on top of a SQLite table. It's not hard.

A local full-text index. The browser already renders the page. At render time, grab the text. Normalize it. Store it in a local SQLite FTS5 table with a budget. When the user hits cmd+F and types a phrase, search across their entire Space, not just the current page. Results come back with the page, the snippet, when you read it, and what branch of your exploration it was in.

Annotations as first-class data. When you highlight a passage, store it against the URL, the selection, a color, an optional note. Don't make it an extension. Make it a database table. It becomes searchable alongside the index. It survives the browser closing. It syncs between sessions because "sync" here means "read the local file again."

None of this needs a model. None of this needs the cloud. None of this needs VC money to run. It needs a team that cares about the problem enough to ship the work.

What we're doing about it

We're building this. It's called Step. It's Chromium underneath, macOS native on top, currently in private beta.

Spaces that actually persist. Everything. Tabs, groups, scroll positions, annotations, split layouts. Close the window, reopen it, nothing is gone.

Step Trail. Your exploration as a visual graph. You can see where you branched. You can drop named checkpoints. You can jump back three days later and land exactly where your thinking was.

Marginalia. Native highlights and notes, five colors, attached to selections, searchable as data. No extension.

Graph Find. Cmd+F across your entire Space. Every page you've visited. Every annotation. Results come with the page, the snippet, the branch, the time.

All local. Your data sits in a SQLite file on your Mac. Your Chrome extensions still work because Chromium still runs underneath. Your passwords carry over. Your bookmarks too.

We're not trying to replace Chrome. We're trying to build what the last thirty years of browser history should have produced if anyone had been paying attention to how people actually work.

If you've ever closed your browser and felt like you lost something, you probably did.

We're building the browser that remembers.