Weave - Private AI journaling that never leaves your phone

Hey PH — solo dev here. Weave is a private journaling app I built so the AI runs entirely on your phone. Nothing you write or say leaves your device. Where it came from: I moved abroad last year for my studies and started having nights where I couldn't sleep. Tried typing a journal — quit after three days because I kept editing myself into the day I wanted to remember. So I switched to just talking into my phone before bed. That stuck. But the first version I built used a cloud AI, and the more I sat with what that meant — every 2am entry going out to a model provider I didn't own — the less I could ignore it. So I spent the last several months rebuilding the whole thing to run on the phone. What's actually running on your phone: Llama 3.2 3B (Q4_K_M) for reflections via llama.rn, Whisper small.en for voice via whisper.rn, MiniLM for semantic embeddings — everything in a local SQLite file that never syncs. Metal GPU offload on iOS. First launch downloads the models over Wi-Fi (~2 GB, once), then the whole app works fully offline, including in airplane mode. How to use it: Open the app — no signup, no account, nothing to configure. Tap the mic and just talk, or type an entry. A few seconds after you save (~11.6s on an iPhone 17), a reflection appears — it either surfaces a pattern from your past entries or asks a question about what you wrote. You can search your journal by feeling, not keywords: type "that thing about my mom last month" and it finds the entry even if you never used the word "mom." Lock it behind Face ID if you want. Honest limits: A 3B model on a phone won't out-nuance a giant cloud model — reflections are decent, not magic. Older phones with 4 GB RAM sit near the memory edge (I unload one model to load the other for a single entry). Android is CPU-only until llama.rn links Vulkan, which makes it slow enough that I haven't shipped it yet — iOS-only for launch. Reflection quality is what I'm most focused on tuning next. Pricing: $3.99/mo with a 7-day free trial, no card needed. No account, no server — the trial is stored locally. That price is basically what the app costs to exist as a one-person, no-data-business. What's next: Better reflections (small-model prompting is a genuinely different craft from big-model prompting), Android once on-device inference catches up, and a couple of features that came out of TestFlight feedback. I'd genuinely love feedback on where it breaks. Especially reflection quality, older-phone performance, and anything in onboarding that felt weird. Reply here or drop a note through the support link in the app.