Vestige: How a Local AI App Tracks Daily Patterns Without the Noise

An ADHD diagnosis often comes with a frustrating paradox: the mind remembers the wrong things at the wrong times. For one developer, the solution wasn’t another productivity app full of prompts or rewards—it was a silent observer that captures spoken moments and surfaces recurring themes over weeks, not minutes. The result, Vestige, is a fully offline Android tracker powered by Google’s Gemma 4 that treats voice entries like raw receipts rather than journal entries.

The app’s core promise is simple: record a 30-second thought, and receive a distilled behavioral pattern—no grades, no emoji reactions, no AI life coaches demanding action. Unlike traditional mood or habit trackers, Vestige avoids subjective scoring entirely. Its goal is to reveal what users already do without realizing it, not to judge or prescribe. For someone who finds daily reflection overwhelming, the absence of performance metrics becomes its own kind of relief.

Built on Gemma 4’s Local Processing

Gemma 4’s on-device architecture was central to the project’s privacy-first design. The developer chose the E4B variant of Gemma 4 specifically because it processes audio input locally without relying on cloud speech recognition services like Google’s SpeechRecognizer. This means voice data never leaves the device—even during transcription.

The workflow unfolds in three stages:

• Capture: A single tap starts recording. Instead of splitting transcription into separate steps, the app streams audio directly into Gemma 4’s inference pipeline.
• Transcription and persona: While still recording, the model returns a transcribed version of the spoken words along with a contextual follow-up response tailored to the speaker’s tone and content.
• Background extraction: After the recording ends, a three-lens convergence process runs in the background, analyzing the entry through literal, inferential, and skeptical lenses before producing a final verdict.

A secondary model, EmbeddingGemma 300M, continuously monitors vocabulary drift. Over time, the system adapts to the user’s evolving language without losing sight of past patterns, ensuring that recurring behaviors aren’t misclassified due to shifting word choices.

Privacy as a Default, Not a Claim

The developer didn’t just promise privacy—they built a verification system to prove it. Every build runs through verifyNoTelemetry, a Gradle task that performs four independent scans to detect any hidden data paths. The results, including privacy receipts, are uploaded as CI artifacts after each run.

Once the Gemma 4 model downloads, the app’s process contains no outbound network connections. The developer demonstrates this by activating airplane mode on camera before recording begins—a visible, reproducible proof that radio signals remain dormant during sensitive operations.

Code structure reinforces the policy:

• NetworkGate seals all networking calls by default.
• No background sync, analytics libraries, or crash reporting tools are included.
• The final APK includes only the essential LiteRT-LM runtime and the Gemma 4 model artifact.

While the developer acknowledges a late-stage oversight—ADRs were numbered but not properly translated into UI-focused acceptance criteria—this transparency underscores the project’s commitment to integrity over marketing fluff.

What Stays Out of the App—and Why

Early prototypes included grid layouts like "Crash," "Deep Space," and "Spiral" to classify moods before speaking. Within days, they were removed. The developer realized that forcing users to label their state upfront defeats the entire purpose: many moments only reveal their shape after they’ve been spoken aloud.

Gamification elements fell just as fast. Gratitude prompts, mood scores, and subscription-based feedback loops were stripped because they implied judgment—something the developer actively avoids. The final version doesn’t grade behavior; it simply stores raw entries as plain markdown files, ensuring users own their data in the most portable format possible.

Even the UI philosophy reflects this restraint. Built with Android’s Jetpack Compose and Material 3, the interface prioritizes clarity over decoration. The developer admits limited mobile experience going in—but the result isn’t a showcase of UI tricks; it’s a functional tool built with deliberate simplicity.

The Code Behind the Silence

The runtime stack is intentionally unremarkable, which was intentional:

implementation("com.litert:litertlm-android:0.11.0")

The project pins the LiteRT-LM library to avoid drift and uses the Gemma 4 E4B model from Hugging Face without shims, wrappers, or parallel inference paths. Audio processing forces CPU usage (AudioBackendChoice.Cpu) due to Gemma 4’s hardware requirements, while text decoding leverages GPU when available.

Core logic lives in functions like ConvergenceResolver, which treats convergence as a pure function—no side effects, no state mutation outside the model’s output. Exports are pulled directly from the ObjectBox database into portable markdown, ensuring entries remain user-owned text files, not locked-in app data.

The entire project is licensed under Polyform Shield 1.0.0, a license designed to prevent commercial exploitation without restricting personal or open-source use.

What Comes Next for Vestige

Vestige isn’t meant to replace therapy or deep reflection—it’s meant to fill a gap in self-awareness where memory and recall lag behind intention. For users who struggle to articulate recurring behaviors in real time, the app offers a quiet alternative to constant prompting.

Future updates may expand pattern recognition without introducing scoring or recommendations. The developer suggests deeper integration with device sensors—accelerometer or ambient light—to enrich context without compromising privacy. One thing won’t change: the refusal to tell users what to do with the patterns uncovered. That part, they say, is already handled.

For now, Vestige stands as a rare example of AI built not to entertain or evaluate, but to observe—then step aside.