A few small improvements, happy to send PRs

[chrisdebian]

Hi, I've been reading through Memoria and like the fully on-device approach. A few things I noticed while building it:

• The README still lists Android 8.0 / API 26 (minSdk is now 28) and describes the old BPE tokenizer and float32 text model, rather than the new SentencePiece encoder.
• There's no CI yet. A basic build-and-lint workflow would catch breakage on PRs.
• Indexing runs in a ViewModel, so it stops if Android kills the app during a large first index. WorkManager or a foreground service might help, but that's your call.

I'd be glad to start with a README fix if that's useful. Do you have any preference on AI-assisted contributions? I use some tooling and would rather check first.

Thanks,

Chris

[raslenabb12]

Thanks for the careful read The README on current master already says API 28 and the int8 text encoder, so please rebase on the latest master before starting. The tokenizer row is outdated, though we moved to a SentencePiece tokenizer in v1.7.0. A README PR for that, the build-from-source file list and the new multilingual feature would be welcome. CI and background indexing are good ideas too, so feel free to open an issue for each before writing code. AI-assisted PRs are fine if you say so in the PR and have tested what you submit.

[chrisdebian]

Thanks for merging the README PR!

I've had a careful read through the code since, and wanted to share a few things I noticed, in case they're useful. I haven't changed any code, and I'll open an issue for anything you'd like me to pick up, as you asked.

1. Search can stop responding after one error. If a search fails once (for example, a shared image that can't be decoded), the results stream ends, and later searches show nothing until the search screen is reopened. The error is only logged, so the user just sees nothing happen. This looks like a small, safe fix.

2. The index can lose photos it can't currently see. When the app scans the gallery, anything in the index that isn't in the scan results is deleted, along with its album memberships. That's right for deleted photos, but it would also happen if someone grants "Select photos" (partial access) on Android 14+, or removes an SD card. I'd like to ask how you'd prefer this to behave before anyone writes code.

3. The released APK lists two permissions the app doesn't use, READ_PHONE_STATE and WRITE_EXTERNAL_STORAGE. They come from the TensorFlow Lite GPU library, which is included but never used, so Android adds them automatically. Removing that dependency should remove both and save about 8 MB.

4. gradlew isn't marked executable in git, so ./gradlew fails from a fresh clone on Linux/macOS (Android Studio is fine). It's a one-line change and would fit inside the CI PR from Add a basic CI workflow (build + lint) for pull requests #9 if you're happy with that.

I also noted a few smaller things (an image-search memory issue, failed photos being skipped silently, new photos not being picked up while the app is open). I'm happy to share those one at a time, whenever suits, rather than all at once.

Chris

[raslenabb12]

Thanks a lot for the careful read, Chris! You can go ahead and fix the search error and remove the unused GPU dependency. One small PR for each, please. I’ll handle the rest myself. Thanks!

[chrisdebian]

Thanks for merging #12 and #13!

I know you said you'd handle the rest yourself, so tell me if these are already on your list. There are three things from the audit I didn't raise last time. I'd be glad to do any of them as a small PR, with an issue first:

1. Indexing gets stuck after a failure. If the indexing worker fails, for example because NNAPI isn't supported on the device, the progress UI never finishes and nothing tells the user why. The fix would show the failure and retry on CPU when NNAPI fails.
2. New photos aren't picked up while the app is open. Photos taken while Memoria is running don't get indexed until it restarts, and the ContentObserver isn't unregistered, so it leaks.
3. A migration test for the Room database. The schema changed several times under version 3. Now that CI is in, I could export the schemas and add a test for upgrading from older versions, so a schema change can't break existing users without anyone noticing.

Happy to do any, all or none of these. Just let me know which.

(I used AI assistance (Claude Code) to review the code, as before. I check everything myself before posting.)