Remember

Remember helps people spatially revisit places held in their existing photographs. It reconstructs only what the photographs can support.

Getting started

1. Open this folder in the latest stable version of Android Studio.
2. Allow Gradle to finish syncing the project.
3. Select an emulator or connected Android device and run the app configuration.

The project uses Kotlin and AndroidX. It targets Android 15 (API 35) and supports Android 7.0 (API 24) and later. The private local Companion currently performs sparse reconstruction; the Android app imports and explores its evidence-backed result.

Documentation

• Product principles
• Reconstruction feasibility spike
• Representation pivot spike
• Semantic planar reconstruction spike
• Photographic coherence spike
• Radiance-field representation spike
• Sparse-view regularization and support-gate spike
• Hybrid supported spatial memory prototype
• Spatial-continuity photo transitions
• Evidence-maximizing structural completion
• Architecture
• Privacy
• Milestone 0 results
• Milestone 0.9 founder steering report
• Milestone 1.0 founder steering report
• Milestone 1.1 founder steering report
• Milestone 1.2 founder steering report
• Milestone 1.3 founder steering report
• Milestone 1.3A founder steering report
• Milestone 1.4 founder review
• Development guide
• Contributing

The Milestone 1.1 private desktop walkthrough is built with tools/milestone09_walkthrough.py. It renders a coarse six-face 3D room shell in pure CSS (no WebGL/Three.js), with pointer-drag look and WASD/arrow translation clamped to the recovered envelope. The walkthrough has a clear Remember view for evidence-backed OBSERVED / RECONSTRUCTED / INFERRED faces, an Imagine view that first attempts validated local learned inpainting on the strongest missing canonical face and otherwise falls back truthfully to the Milestone 1.0 deterministic IMAGINED structural atlas fallback, and a debug provenance view. Completion happens in canonical room-face/atlas space, never frame-by-frame; OBSERVED/RECONSTRUCTED, non-generatable, and critical regions are restored/locked exactly, ambiguous/critical unknowns remain ABSENT, and object gap completion is disabled unless confidence is adequate. The learned path is local-only, fixed-seed, fail-closed, and only becomes the default when the configured Diffusers inpainting model is available and validates. The tool emits aggregate metrics and provenance-coded views in the requested output directory; private photos, poses, masks, models, and renders must remain outside the repository.

Milestone 1.4 freezes an already accepted private walkthrough without invoking reconstruction or generation. tools/milestone14_demo_freeze.py accepts only the exact approved front/back/right asset hashes and the accepted 1.3/1.3A metrics, requires an output directory outside Git, and emits the minimal offline presentation bundle plus a deterministic 36-second backup video. tools/milestone14_synthetic_test.py validates the freeze contract without reading private data.

Licensing

Copyright (c) 2026 Remember contributors. All rights reserved.

This repository is public for development visibility. No license is granted to use, copy, modify, or distribute its contents.