Memora

Ask anything about your data.

Desktop app that indexes your local files (PDF, DOCX, text, markdown, code) and lets you ask natural-language questions. Answers come back with citations that open the original file.

Built for developers and tech workers. Privacy-first: local-default, cloud opt-in.

See PLAN.md for the full V1 implementation plan.

Stack

• App: Electron + TypeScript + React (Vite)
• Vector DB: LanceDB (embedded)
• Metadata: better-sqlite3
• Embedding: BGE-M3 via onnxruntime-node
• Reranker: BGE reranker base via onnxruntime-node
• LLM: node-llama-cpp (bundled) with Ollama auto-detect
• Search: hybrid (BM25 + vector + reciprocal rank fusion) + cross-encoder rerank
• File watch: chokidar

Monorepo layout

apps/
  desktop/            Electron app (main + preload + renderer)
  context-menu/
    macos/            Finder Sync Extension (Swift, Xcode) — Phase 3
    windows/          Shell extension (C# COM) — Phase 3
packages/
  shared/             Types shared across workspaces
  eval/                Golden set runner + RAGAS client
services/
  ragas-eval/         Python RAGAS eval API (docker-compose)
resources/
  models/             BGE-M3 + reranker ONNX (fetched at runtime, .gitignored)
  binaries/           Platform-native binaries (built in CI, .gitignored)

Prerequisites

• Node.js >= 20.11
• Docker Desktop (for dev services only — the Electron app runs natively)
• macOS or Windows 11 (Linux supported for dev only; no context-menu integration)

Setup

# 1. Install npm deps
npm install

# 2. Start dev services (Ollama + RAGAS eval runner)
npm run services:up

# 3. Run the Electron app in dev mode
npm run dev

What docker-compose runs

Docker is used only for dev-time backing services. The Electron app itself runs natively (it needs display access + native OS APIs; Electron cannot run in a container usefully).

Service	Port	Purpose
ollama	11434	Dev LLM server. App auto-detects and uses this when available.
ollama-bootstrap	–	One-shot container that pulls qwen2.5:7b-instruct + bge-m3.
ragas-eval	7860	Python FastAPI wrapper around RAGAS for search-quality evaluation.

Stop services: npm run services:down. Logs: npm run services:logs.

Production runtime

In production, the app ships node-llama-cpp with a bundled Qwen2.5-7B-Instruct Q4 model, downloaded on first launch. It still detects Ollama at localhost:11434 and defers to it when the user has their own Ollama running (respects power-user setup with larger models).

Development commands

npm run dev              # Run Electron app (main + renderer with HMR)
npm run build            # Build all workspaces
npm run typecheck        # Type-check all workspaces
npm run lint             # ESLint
npm run test             # Vitest
npm run services:up      # Start Ollama + RAGAS eval
npm run services:down    # Stop dev services

Roadmap

Phase 0 (Week 1): scaffolding ← you are here Phase 1 (Week 2–5): indexing pipeline (local FS → parser → chunker → embedder → LanceDB) Phase 2 (Week 6–8): query engine + UI (hybrid search → rerank → LLM synthesis, chat UI, citations) Phase 3 (Week 9–10): context-menu integration (macOS Finder Sync + Windows shell ext) Phase 4 (Week 11): eval framework, encryption at rest, code sign, crash reporting Phase 5 (Week 12): beta launch

See PLAN.md §3 for full breakdown.

Privacy

• Local default: no network calls in local mode. Enforced via renderer CSP and main-process opt-in flag.
• Cloud opt-in (v1.5+): explicit toggle with warning modal before any query is sent off-device.
• Data at rest: LanceDB + SQLite encrypted with a key stored in the OS keychain (macOS Keychain / Windows Credential Manager).
• No telemetry by default: crash reports opt-in with PII scrubbing.

See PLAN.md §7 for the full privacy checklist.