Thought Bridges: How ThoughtTrack Connects Your Ideas Without a Server

The Problem No One Talks About

Voice note apps are good at capture. They're terrible at recall.

You speak 200 thoughts over six weeks. Now you need the one about the project structure you mentioned on a Tuesday, somewhere between the coffee shop recommendation and the birthday gift idea. Good luck finding it.

The usual answer is search — a text field that matches keywords. That helps, but it misses the harder problem: most related thoughts don't share keywords. The idea you had about your team's communication style in November might be deeply relevant to the friction you're feeling in a project meeting in February. Search won't connect those. They don't use the same words.

This is what Thought Bridges is for.

What Thought Bridges Does

Thought Bridges is a feature in ThoughtTrack that surfaces semantic relationships between your thoughts. When you open a thought, the app shows you other thoughts it's meaningfully connected to — not because they share keywords, but because they share meaning.

Finding the right word for what you half-remember isn't the goal. ThoughtTrack finds the connections you didn't know to look for.

How It Works

Thought Bridges uses a composite scoring system combining three signals:

Semantic embeddings (55% weight) are the core. Apple's NaturalLanguage framework includes `NLEmbedding`, which generates a numerical vector representation of any piece of text. Two thoughts with similar embeddings are semantically similar, even if they use completely different words. "I'm worried about how we're coordinating" and "the left hand doesn't know what the right hand is doing" will score as related — because they mean the same thing.

Named entity recognition (25% weight) uses Apple's NER to extract people, places, and organizations from your thoughts. If two thoughts both mention a specific person or project name, that shared reference is a strong signal they belong together — even if the surrounding text is different.

Tag overlap (20% weight) catches cases where you've manually tagged thoughts. Tags are user-defined and explicit, so they're a high-confidence signal when they match, but they're absent for most thoughts. This is why tag overlap is weighted lowest — it's useful but sparse.

When the composite score exceeds a threshold, ThoughtTrack creates a bridge between the two thoughts. Low-confidence connections are filtered out. The bridges you see are the ones the model is reasonably confident about.

Why It Runs On-Device

This was never a design choice we debated. It was a constraint we embraced before writing the first line of code.

Your thoughts are not a product feature. They're not training data. They're not anonymized signals. They are the most intimate content you produce — more revealing than your photos, more personal than your messages. Sending them to a server for processing, even a server we control and trust, is a line we won't cross.

Apple's NaturalLanguage framework makes on-device NLP practical. `NLEmbedding` runs locally. NER runs locally. The whole scoring pipeline runs on your phone, on your data, and the output never leaves your device. We don't see it. We can't.

This has a cost. On-device models are smaller than server-side ones. The embeddings are less nuanced than what a large language model would produce. The NER is less sophisticated than cloud APIs. We accepted those tradeoffs.

The connections Thought Bridges surfaces are genuinely useful and sometimes surprising. We tested thoroughly with realistic thought datasets. The false positive rate is low. The bridges that appear are real.

What It Looks Like in Practice

Open any thought in ThoughtTrack. If there are related thoughts, they appear in a "Bridges" section below the main content — each with a short excerpt and when you captured it.

A common pattern from beta testing: a user captures a thought about feeling stuck on a decision, and Thought Bridges surfaces a thought from two months earlier where they captured a similar stuck feeling — along with the resolution they eventually found. They'd completely forgotten that earlier thought. Seeing it answered the current question.

That's the interaction we were aiming for. Not clever technology for its own sake, but a moment of "oh, right."

---

ThoughtTrack launches on the App Store March 2026. Join the waitlist for a launch-day notification.