Mind-Forge Pro

An offline, on-device "second brain" AI agent for Android. No Termux, no external Ollama server — the app embeds llama.cpp natively via Dart FFI and serves itself over a local HTTP API.

How it works

• Chat — talk to a locally-loaded .gguf model. Every turn is grounded with a lightweight TF-IDF retrieval pass over your own stored notes (lib/memory_agent.dart) before it reaches the model.
• Memory DB — browse, inspect, and delete anything the app has remembered (SQLite, on-device only).
• Server Config — pick a .gguf file from local storage and start a local API server (http://127.0.0.1:<port>/api/generate) that any other app or script on the device can call.

Architecture

Layer	Tech
UI	Flutter (Material 3)
Inference	llama_cpp_dart, off the UI thread via LlamaParent/LlamaScope
Local server	shelf + shelf_router
Memory / RAG	sqflite + a pure-Dart TF-IDF keyword ranker
Native actions	a MethodChannel the model can trigger via <ACTION: X> tags

Building

CI (.github/workflows/build-apk.yml) does everything needed for a sideloadable release APK — no production keystore required:

1. Compiles libllama.so (+ its libggml*.so dependencies) from the llama.cpp source via the Android NDK, since llama_cpp_dart 0.2.x is a pure FFI binding and ships no prebuilt binary itself. The result lands in android/app/src/main/jniLibs/arm64-v8a/.
2. Runs flutter build apk --release --target-platform android-arm64 with R8/ProGuard shrinking and native debug-symbol stripping enabled (android/app/build.gradle).
3. Signs the release build with the standard Android debug keystore, cached across runs so every build shares the same signing key — that's what lets you install an updated APK over a previous one without uninstalling first. (This is fine for sideloading; it is not a Play Store–ready signature.)
4. Uploads the finished APK as a workflow artifact.

To build locally instead, you'll need the Android NDK on your PATH and a compiled libllama.so for your target ABI in android/app/src/main/jniLibs/<abi>/ before running flutter build apk.

Status

Early scaffold — chat, memory, and server-management UI are wired up; model loading and generation go through the real llama_cpp_dart API (LlamaParent + LlamaScope), but this hasn't yet been run against a real device/model. Treat native-library and API surface details as "verify against whatever llama_cpp_dart version is pinned in pubspec.yaml" — that package's public API has changed release to release.