MemRE

AI-powered spaced repetition flashcards as a native macOS desktop app.

---

Features

• 🃏 Spaced repetition — SM-2 inspired scheduler with continuous retention curve
• 🧠 Multi-provider AI — Ollama (Gemma 4), Anthropic API (Claude Sonnet), Claude Code CLI, deterministic offline stub fallback
• 🎚️ Per-call provider override — pin Sonnet for one card-generation pass without changing your default
• ♻️ Hot-reload settings — switch providers / paste an API key, no restart
• 📚 Knowledge sources — paste a study text and have AI extract cards from it (FractoP-comprehensive mode for long content)
• 🔌 Local-first — SQLite at ~/Library/Application Support/Memre/memre.db. API keys never leave the machine
• 🪟 Native macOS — single 15 MB .app bundle, no Electron, no Chromium

Install

Download the latest signed .app from Releases:

# Apple Silicon
curl -L -o MemRE.zip https://github.com/O6lvl4/memre/releases/latest/download/memre-v0.1.3-darwin-arm64.zip
unzip MemRE.zip
mv memre.app /Applications/
xattr -dr com.apple.quarantine /Applications/memre.app   # ad-hoc signed only
open /Applications/memre.app

Then click the gear icon in the header to choose an AI provider.

AI providers

Provider	Auth	Default model	Cost
Ollama	none — local daemon	gemma4:26b (or any tag installed)	free, runs on your machine
Anthropic API	API key	claude-sonnet-4-6	per-token (your account)
Claude Code CLI	none — uses existing claude session	claude-sonnet-4-6	included in your Claude subscription
Local stub	n/a	offline fallback	free, low quality (sentence splitter)

Settings are stored locally in SQLite. The Anthropic key is plain text in the same DB; never sent anywhere except api.anthropic.com. Claude Code provider shells out to your claude binary so authentication is delegated to that.

Architecture

internal/
├── deck/           bounded context: entity + service + repo + sqlite + Wails handler
├── card/
├── knowledge/
├── ai/             Provider port + Ollama / Anthropic / ClaudeCode / Stub + Registry
├── settings/       KV store, hot-reload backend
├── srs/            pure SM-2 / retention math
├── platform/
│   ├── sqlite/     shared connection + idempotent migrations
│   ├── clock/      Clock interface + System + Fake (for tests)
│   └── idgen/      Generator interface + Crypto + Sequential
└── composition/    composition root — only place adapters meet ports

2026 Go DDD style: vertical slice (one bounded context per directory), consumer-side Repository interfaces, Service structs with methods, no separate domain/application/infrastructure/ layer dirs. Tests substitute fakes in-package without any DI framework.

Context map

_{Source: docs/assets/context-map.mmd — rendered with premaid (pretty theme).}

Bounded context relationships

From → To	Relationship	Notes
deck ← card data	read model via StatsRepository	deck never imports card; deck-level stats come from a dedicated read-model port computed against the cards table
card → events	published events	card.Reviewed lets future gamification / achievement subscribers react without coupling
card / deck → srs	shared kernel	pure scheduling math used by both, no state
ai → external LLMs	anti-corruption layer	Provider interface insulates the rest of the app from JSON Schemas, model names, and provider quirks
every slice → platform	infrastructure	each slice owns its own Sqlite* adapter that depends on the shared Store
composition → all	composition root	the only place concrete adapters meet ports

Quick start (development)

Requirements: Go 1.25+, Node 20+, Xcode CLT, wails3 CLI.

git clone https://github.com/O6lvl4/memre
cd memre

# install deps + generate Wails TS bindings
cd frontend && npm ci && cd ..
wails3 generate bindings -ts

# run with hot reload
wails3 dev

# or build a production .app bundle
wails3 task darwin:package
open bin/memre.app

Tests

go test -race -count=1 -cover ./internal/...

Package	Coverage	Notes
srs	90%	pure SM-2 + retention math, table-driven
deck	85%	entity invariants + service + sqlite roundtrip
knowledge	73%	+ ON DELETE CASCADE, FK enforcement
card	66%	+ ApplyReview, IsDue, upsert
ai	52%	Ollama/Anthropic via httptest, Registry resolve, Fallback
settings	60%	KV roundtrip, upsert
platform/clock	100%
platform/idgen	94%
platform/sqlite	36%	migration + WAL/foreign_keys pragmas + idempotency

≈110 unit + integration tests, all run on every CI pipeline.

Custom providers / models

# pick a different Ollama model
MEMRE_OLLAMA_MODEL=gemma4:e4b open bin/memre.app

# point at a remote Ollama
MEMRE_OLLAMA_URL=http://192.168.1.50:11434 open bin/memre.app

# or set defaults via the in-app gear icon (persists in SQLite)

Releasing

A git push of an annotated tag vX.Y.Z triggers .github/workflows/release.yml, which builds on macos-latest, ad-hoc signs the .app, zips it, computes SHA-256, and attaches both to a GitHub Release with auto-generated notes.

git tag -a v0.2.0 -m "..."
git push origin v0.2.0

CI

.github/workflows/ci.yml:

• Linux: go vet + go test -race -cover on the pure layer (no Wails GUI deps)
• Linux: frontend tsc -b && vite build after generating TS bindings
• macOS: full graph go vet ./... + .app smoke test + binary artifact upload

Roadmap

• Streaming AI responses (token-by-token rendering for long generations)
• macOS Keychain for the Anthropic API key (currently SQLite plaintext)
• Apple Developer ID signing + notarisation (Gatekeeper friction-free install)
• Universal binary (arm64 + amd64) and Linux / Windows builds
• Auto-update (check GitHub Releases on launch)
• Import / export Anki .apkg and CSV
• Slog structured logging + per-call AI metrics

License

MIT © 2026 MemRE contributors