MemoryVault

A self-hosted video archive with automatic transcription and a semantic memory map.

Services

Service	Directory	Port
API	backend/	8000
Worker	worker/	—
Frontend	frontend/	8765

Quick start

cp .env.example .env
# Set MEDIA_DIR to your video folder in .env
docker compose up --build

Frontend: http://localhost:8765
API: http://localhost:8000

How it works

1. Drop video files into $MEDIA_DIR/new/
2. Hit Scan in the UI (or POST /scan) — the API creates job records
3. The worker picks them up, remuxes MTS files, transcribes with WhisperX, and writes results to the shared SQLite database
4. Use the semantic map to browse videos by content similarity

Shared database

All services share a single SQLite file at ./database/app.db (WAL mode, safe for concurrent reads + one writer).

Configuration

.env

Variable	Service	Default	Description
MEDIA_DIR	compose	—	Required. Path on the host to your video archive
MEMORYVAULT_BASE	both	/data	Container-side mount point for the video archive
THUMBNAILS_DIR	both	/app/thumbnails	Container-side path for thumbnails
DATABASE_URL	both	sqlite:////app/database/app.db	SQLAlchemy DB URL
LOG_LEVEL	both	INFO	Log verbosity (DEBUG, INFO, WARNING, ERROR)
OLLAMA_URL	backend	—	Base URL of your Ollama instance (e.g. http://192.168.1.10:11434)
OLLAMA_MODEL	backend	qwen3:8b	Ollama model name to use
OLLAMA_TIMEOUT_SEC	backend	120	Request timeout in seconds
EMBEDDING_MODEL	backend	microsoft/harrier-oss-v1-0.6b	Embedding model for semantic search

Ollama

MemoryVault uses Ollama for LLM-powered features. Set OLLAMA_URL in .env to point at any reachable Ollama instance — a machine on your local network, a remote server, or a local process:

# Remote machine on the LAN
OLLAMA_URL=http://192.168.1.10:11434

# Ollama running directly on the host (not in Docker)
OLLAMA_URL=http://host.docker.internal:11434

Make sure the model named in OLLAMA_MODEL has already been pulled on that machine (ollama pull qwen3:8b).

Service-specific variables

These are set directly in compose.yml and rarely need changing:

Variable	Service	Default	Description
CORS_ORIGINS	backend	http://localhost:3000,...	Comma-separated allowed origins
WORKER_POLL_INTERVAL	worker	5	Seconds between DB polls for picking up background jobs

GPU acceleration

Change TORCH_INDEX_URL in compose.yml to https://download.pytorch.org/whl/cu124 for CUDA 12.4.