OpenRecall

An Arabic, dark, keyboard-accessible desktop flashcard app for personal spaced repetition. Built with Tauri 2, React/TypeScript, SQLite and a native FSRS-7 engine. Windows 11 x64 is the primary target. Releases also provide Windows 10 x64 and Linux x64 packages. macOS downloads remain available in releases through 0.2.0.

Download releases · Arabic usage guide · Architecture · Contributing

Study workflow

• One optional spelling deck collects unique Latin-script words/terms from an ordinary source, including old cards and every alternative form. Source edits, deletion and rebinding retain collected progress and last matching contexts. Explicit word deletion creates an exclusion; allowing collection again uses the current source with fresh progress.

• Spelling review prefills the target on the native textarea's first line, with focus and a collapsed caret at its beginning and a blank line below. Typing is uncontrolled and disposable: leaving the presentation destroys every user edit, including on undo or restart. The application never reads/copies attempts, checks spelling or grades automatically. The four grades are available immediately.

• Per-deck and collapsed library statistics show daily completions based on the last effective review per logical card and configured study day, with a separate spelling-word subset. Original-card study and spelling are independent tasks.

• Optional per-form questionTitle customizes the label-only H1. Missing/null or blank labels use السؤال; omitted variant labels inherit the root on JSON import, while explicit null/blank resets them. Labels are plain text, limited to 160 Unicode characters without control characters or newlines.

• Create, rename and delete decks. Each deck shows logical card counts, due/new cards, review history, forecasts and scheduling settings.

• Create/edit/delete cards, including any number of alternative wordings. All forms belong to one logical card and share its memory state and due date.

• Forms are shuffled without replacement through a persistent bag. With more than one form, the boundary between cycles avoids an immediate repeat. Revealing or switching filters does not consume a form; grading does.

• Start with all due cards, including new and due learning cards, ordered by due time. Filter to never-reviewed or previously-reviewed cards. Cards becoming due during an active session join its queue without interrupting the current question.

• Reveal the answer, then choose Again/Hard/Good/Easy using Arabic labels. Each rating's accessible name includes its approximate interval; the committed result shows the actual next due time. Undo restores the previous question, answer, learning state, form bag, queue and session counters.

• Session summaries distinguish attempts, unique cards, ratings and active study time. Review headings are the labels السؤال (H1), الإجابة (H2) and الملاحظات (H3 when notes exist). Card content remains ordinary text.

Statistics, card lists and advanced controls start collapsed. The interface uses Arabic RTL navigation and dir="auto" for multilingual card content. Noto Sans Arabic is bundled locally. Text size, reduced motion and review shortcuts are configurable.

Native scheduling and optimization

The engine is pinned to upstream commit 0fd59949e9168af7194418c2dba99836b5548e87. That revision contains the 34-parameter FSRS-7 model, including stability_fast, although its crate package version still says 6.6.2. The published registry crate is not substituted based on a version number alone. Numerical scheduling goldens, intraday memory and the real optimizer are tested in Rust.

Named optimization profiles can serve one or several selected decks. Detaching a deck preserves its current model and creates independent future training. Training uses only the explicit membership of that profile. Schedule penalties default on. Minimum data requirements include 400 usable nonfirst observations, 50 cards, seven study days, both recall outcomes, and a chronological held-out set of at least 80 observations from 10 cards. Candidates require improved held-out log loss with bounded RMSE regression and explicit application. Changed history, membership, configuration or baseline invalidates stale candidates.

Defaults: Cairo timezone, study day beginning at 01:00 local time, retention 90%, minimum interval 60 seconds and maximum interval 36,500 days. Advanced fields and per-deck overrides have validated bounds. Changing a model or policy preserves already saved due dates; subsequent reviews use the new model/policy.

JSON import and complete backups

Card import accepts UTF-8 JSON, including a BOM, up to 50 MiB:

[
  {
    "questionTitle": "عرّف",
    "front": "ما معنى الاسترجاع النشط؟",
    "back": "محاولة استحضار المعلومة قبل رؤية الإجابة.",
    "notes": "أجب أولًا ثم اعرض الإجابة.",
    "variants": [
      {
        "questionTitle": "اشرح",
        "front": "اشرح الاسترجاع النشط بكلماتك.",
        "back": "استحضار المعلومة من الذاكرة قبل الاطلاع عليها.",
        "notes": ""
      }
    ]
  }
]

variants can be absent or empty. Preview shows invalid rows/forms and duplicate matches before any library write. Select valid cards and explicitly exclude bad forms. Duplicate identity uses NFC-normalized, whitespace-collapsed question and answer across ordinary decks; case, punctuation and diacritics remain significant. Managed spelling tasks do not exclude ordinary imports. Notes and question labels do not create a distinct card identity. Import commits atomically against the preview's library revision.

Complete backups are a separate versioned UTF-8 JSON format, currently version 2 with a shared 256 MiB export/restore limit. They include settings, decks, forms, progress, review/action/undo history, optimization profiles, model provenance and training reports, spelling vocabulary, retained source contexts and exclusions. The reader also supports genuine v0.1.0 format-1 exports and verifies their original payload digest before upgrading defaults. Backups are plaintext, with a canonical payload digest for corruption detection. Export finalizes a collision-safe file in Downloads, or opens a native save dialog if that location is unavailable. Restore previews its counts, requires full-replacement confirmation and writes a safety backup first. Invalid, corrupted or future-format data does not replace the current library.

Installation, offline use and updates

The normal Windows NSIS installer uses an existing WebView2 runtime or downloads it during initial setup. The larger offline-setup.exe embeds the Evergreen WebView2 installer for initial setup without an Internet connection. SQLite, fonts and the frontend are bundled. End users do not install Node, Rust or a database server. Studying, editing, importing, backups, statistics and training run locally without network access.

Updates are requested in Settings: check, download, then explicitly confirm installation. The official Tauri updater verifies package bytes and their signed version. Installation waits for saved edits, completed/ended review and durable training state. See signing and release operations. Updater signing is separate from OS publisher certificates. Windows packages have no Authenticode publisher identity.

Use the Linux AppImage for in-app updates; .deb is additional installation media managed through the system package installer. Linux builds use Ubuntu 22.04 and WebKitGTK 4.1. Secondary-platform builds and actual runtime testing are reported separately in verification results.

Development and verification

Use Node 24.21, pnpm 12.15 and Rust 1.99 with MSVC Build Tools on Windows or the Tauri native prerequisites on macOS/Linux. See toolchain decisions.

pnpm install --frozen-lockfile
pnpm tauri dev
pnpm verify

pnpm verify runs Rust formatting, strict Clippy, workspace tests, generated DTO drift, TypeScript, ESLint, frontend formatting/tests, maintenance policy tests, the frontend production build and workflow linting. Browser/native QA commands and their environment are documented with the verification results. A checksum- verified actionlint binary is installed under ignored target/tools if needed.

GitHub requires exact-revision quality/native/platform checks. Dependabot opens dependency PRs; only verified same-repository development patch/minor changes within the tested policy can merge automatically. Runtime, Rust, Tauri, FSRS, workflow and security changes require review. A quiet, deduplicated FSRS watcher tests a newly published crate against the actual core scheduling/optimizer tests on an isolated runner before opening a migration issue. Release jobs compile without signing credentials, sign existing bundles, verify all artifacts and publish one complete updater manifest.

Agent accessibility testing uses automated DOM/keyboard/focus/axe checks. NVDA testing belongs exclusively to the human partner. The agent must never launch, control or configure NVDA, Narrator or another screen reader. The manual checklist is delivered independently of source/installer completion.

License

OpenRecall code is MIT licensed. Bundled libraries/fonts retain their own licenses; see third-party notices and the complete license texts under licenses/, also included in installed packages.