Heyday Alternative: Why I Built a Local-First Web Assistant

Heyday Alternative: Why I Built a Local-First Web Assistant

Here's something most people in the productivity tool space won't say plainly: any tool that uploads a continuous record of everything you read, search, and click to someone else's servers is doing something fundamentally reckless, no matter how pretty the UI is.

I don't care if the company pinky-promises they encrypt it. I don't care if their privacy policy is written in friendly, approachable language. The moment your ambient browsing behavior leaves your machine and lands on infrastructure you don't control, you've created a liability that didn't need to exist. And Heyday, for all its genuine polish, asks you to do exactly that.

I've been using TraceMind as my daily driver for about six months now. Before that, I spent a few weeks with Heyday. Before that, I was duct-taping together a mess of bookmarks, Notion databases, and the naive hope that Chrome's built-in history would actually help me find something from two weeks ago. (It won't. I wrote about why Chrome's history falls short if you want the details.)

This isn't a hit piece. Heyday does some things well. But I think their architectural choice, cloud-first storage of browsing behavior, is a mistake that matters more than most users realize. And I think there's a better way to build this kind of tool.

What Heyday actually does well

Credit where it's due. Heyday's resurfacing concept is clever. You're writing a doc, and it quietly shows you pages you visited weeks ago that might be relevant. The idea of a "memory assistant" that watches what you browse and then reminds you of forgotten context is legitimately useful. I felt the value of it within days.

Their UI is clean. Onboarding is smooth. The search works reasonably well for keyword matching. If I were evaluating Heyday purely on "does it solve a real problem," the answer is yes, absolutely.

But solving a real problem doesn't excuse solving it dangerously.

The part that kept me up at night

When I was using Heyday, I started paying closer attention to what it was actually collecting. And the answer is: basically everything. Every page I visited. Content from those pages. My browsing patterns. My reading habits. The things I searched for.

All of it, sent to their servers.

Think about what that dataset looks like after six months. It's a comprehensive map of your professional interests, your personal curiosities, your health concerns, your political leanings, what you're researching at 2 AM when you can't sleep. It's arguably more intimate than your email.

And it's sitting in someone else's database. Subject to their security practices. Their employee access policies. Their response to government data requests. Their future acquisition by a company with different values.

I keep coming back to a simple principle: ambient behavioral data, the passive record of everything you do, is categorically different from data you actively choose to share. When I post something on Twitter, I'm choosing to make it public. When a tool silently catalogs every article I read about a medical symptom, that's something else entirely. That data should never leave my device unless I explicitly, consciously decide to send it somewhere.

Why "we encrypt it" isn't enough

I can already hear the counterargument. "But they encrypt the data. It's secure."

Three problems with this.

1. Encryption protects data in transit and at rest from outside attackers. It does nothing about insider access, government subpoenas, or the company itself deciding to use your data differently after a policy change.
2. You're trusting their implementation. Have you audited it? Has anyone? When a company says "encrypted," that word is doing a lot of heavy lifting with very little accountability.
3. Even with perfect encryption, the existence of a centralized database of millions of users' complete browsing histories is a target. You've created a honeypot. The question isn't whether someone will try to breach it, it's when.

The genuinely safe approach is to never create that centralized target in the first place.

What "local-first" actually means in practice

So what's the alternative? This is where I should talk about TraceMind, since I've been living inside it daily.

TraceMind is a Chrome extension that captures readable text from eligible pages, similar in concept to Heyday. The critical architectural difference is that the searchable corpus, embeddings, and core search stay on the device. Licensing calls TraceMind's API. Optional Pro Chat has its own explicit boundary: it sends the question, selected excerpts, titles, URLs, and current chat context directly to the configured provider when invoked.

The core captured corpus lives in IndexedDB, right in your browser. ML inference (more on that in a second) prefers WebGPU and falls back to WASM. Captured readable text, embeddings, notes, and tags stay on the physical device in front of you unless selected source context is deliberately sent through Pro Chat.

This isn't a philosophical stance slapped onto a product as marketing. It's a genuine architectural decision that shapes everything about how the tool works. And honestly, it introduces constraints that cloud-based tools don't have. I'll get to those.

The search is the thing that sold me

Here's where it gets interesting. Heyday's search is keyword-based, which means you need to remember specific words from the page you're looking for. That works sometimes. Often it doesn't.

TraceMind runs a model called all-MiniLM-L6-v2, a 384-dimensional embedding model, directly in your browser via WebGPU (or WASM as a fallback). What that means in practice: you can search by meaning, not just by exact words.

Last month I was trying to find an article I'd read about companies quietly reducing package sizes instead of raising prices. I couldn't remember what it was called. I searched "shrinkflation consumer goods" in TraceMind, and it surfaced the exact article, even though the word "shrinkflation" didn't appear anywhere in the piece. The article had used phrases like "reducing product volume" and "stealth price increases." The semantic search connected the dots.

That moment, more than anything else, is what made me stop looking for alternatives.

TraceMind also combines this semantic search with traditional full-text search (using FlexSearch) and merges the results through something called Reciprocal Rank Fusion. So if you do remember the exact phrase, that still works great. You get both approaches at once. I've written more about how semantic search differs from keyword search if the technical side interests you.

Tradeoffs I won't pretend don't exist

Being honest. TraceMind isn't magic, and the local-first approach has real costs.

First: it only knows about pages captured in the browser where it is installed. There is no automatic cross-device sync. You can move data manually with a passphrase-encrypted backup; creating a new encrypted backup requires Pro, while importing or decrypting one is free. New passphrases use 600,000 PBKDF2 iterations, and legacy 200,000-iteration data remains readable.

Second: the initial indexing of a page takes a moment. The ML model is running on your hardware, not a beefy cloud GPU. On my M2 MacBook Air it's fast enough that I never notice. On an older machine? Could be different. I haven't tested extensively on lower-end hardware.

Third: storage. All this lives in your browser. If you clear your browser data aggressively, you could nuke your TraceMind database. (The content is compressed 50-70% with lz-string, so it's not as bloated as you'd expect, but it's still local storage you need to be aware of.)

These are real tradeoffs. I think they're worth it. You might not. That's fine.

The stuff I actually use daily

Let me get concrete about my workflow, because abstract privacy arguments only go so far.

Every morning I open TraceMind's search and type something vague like "that API pricing comparison from last week." It can find an eligible page without a tag or bookmark because it indexed the readable text when I visited and can match "API pricing comparison" semantically.

I use the Offline Page Viewer constantly. It saves sandboxed HTML reading copies. They can remain useful if an article changes or becomes unavailable, but they are not guaranteed complete archives; scripts are removed and images or other resources depend on what capture succeeded. This is a Pro feature.

The tagging system with AI suggestions is useful but not life-changing. Honestly, I use it maybe twice a week. What I really rely on is search. The tags are gravy.

One thing that surprised me: the analytics. TraceMind has a drill-down analytics view that shows you your browsing patterns. I discovered I was spending roughly 40 minutes a day on documentation sites, which made me realize I should probably just set up a local docs mirror for the frameworks I use most. Not the flashiest feature, but genuinely useful for understanding where my time goes.

The free tier is actually usable

This matters. TraceMind's free tier includes the full semantic and keyword search engine, low-quality screenshot capture, up to 10 excluded domains, no TraceMind page cap, and keep-forever captured-history retention subject to browser storage capacity. Free users delete history manually; Pro can choose a finite automatic window.

It's not a crippled demo. The free tier includes the full semantic and keyword search engine. Pro adds the Offline Page Viewer, screenshots up to 1920x1080, notes, optional passphrase encryption for supported local content and new encrypted backups, and unlimited domain exclusions.

A direct comparison, since that's why you're here

If you want a detailed side-by-side breakdown, there's a thorough TraceMind vs Heyday comparison that covers features point by point. But here's my opinionated summary:

Heyday wins on: multi-device sync, automatic resurfacing in Google Docs, overall UI polish, the "it just works in the background" feeling.

TraceMind wins on: privacy (by a mile), search quality (semantic vs keyword), offline access to saved pages, and the fact that your browsing history isn't funding someone else's business model.

Pick which list matters more to you. For me, it wasn't close.

Why I think local-first is the only ethical default

I want to be careful here because I don't think the people building Heyday are bad actors. I think they built a useful product with the architecture that was easiest and most natural for a cloud-based startup. Upload data, process it on servers, deliver results. That's how most software works.

But browsing history isn't "most data." It's a comprehensive behavioral record. It reveals what you think about, worry about, and are curious about, with timestamps. The default for that kind of data should be: it stays on your device. Period. If you want to sync it somewhere, that should be an opt-in choice you make with full understanding.

TraceMind takes that position as a core engineering constraint, not a feature checkbox. The core captured corpus stays local, ML runs locally, and search happens on your machine. Licensing contacts TraceMind. Optional Pro Chat sends selected sources and the current conversation directly to the configured provider. You can inspect those boundaries in Chrome DevTools rather than relying on a blanket "no network" claim.

I've watched enough "we take your privacy seriously" companies get acquired, breached, or quietly update their terms of service that I don't think caution here is unwarranted. If you want to read more about on-device AI and why it matters for privacy, I've gone deeper on the technical side elsewhere.

The bottom line

Heyday identified a real problem: our brains can't remember everything we read online, and browser history is garbage at helping us find it again. Full agreement there.

Their solution asks you to hand over a detailed behavioral record to a third party. TraceMind solves the retrieval problem with a local core; the only browsing context it sends externally is the selected material and conversation a user deliberately supplies to optional Pro Chat.

Six months in, I've indexed thousands of pages. I find things I'd forgotten I read. My browsing history is genuinely useful for the first time, and the searchable corpus remains in the browser on the laptop in front of me.

That's the tool I wanted. So that's the one I use.