Aura is a self-hosted AI agent platform with temporal graph memory
Summary
Aura is a provider-neutral, self-hosted AI agent platform written in Go for ongoing work on infrastructure controlled by its operator. Its runtime includes a CLI, web cockpit, Telegram gateway, tools, document retrieval, scheduled jobs, and a temporal graph memory that stores facts, sources, and validity windows in ArcadeDB; conversations are authoritative in Postgres. The platform supports multiple users, giving each identity its own memory database, workspace, sandbox, capability grants, and audit view. Models can be selected in the cockpit through OpenRouter, a ChatGPT plan, the bundled llama.cpp server, or Ollama; the default route uses DeepSeek-V4 Flash through OpenRouter. Aura can run inference through a cloud provider or a local server, but its local storage does not by itself make cloud inference offline because sent context still reaches the selected provider. Its agent loop includes shared budgets, repeated-call controls, deferred tools, tool discovery, adaptive reasoning, terminal and filesystem access, approval gates, secret redaction, and MCP server connections. Document retrieval provides indexed passages, source hashes, citations, and access to original files, while a per-identity sandbox can run with gVisor on native Linux when the optional sandbox profile is enabled. A scheduler supports one-time, recurring, and cron jobs, including full agent jobs whose outcomes return to the owning conversation. The cockpit also provides voice input and output, mid-turn steering, approvals, image and video generation through OpenRouter media models, browser-based editing, and a multi-track video editor. Telegram, WhatsApp, calendar and email integrations, web search through bundled SearXNG, and Cloudflare remote access are supplied as integrations, with WhatsApp using an unofficial client that carries account-ban and terms-of-service risk. Aura is distributed as a Docker Compose appliance with Postgres, ArcadeDB, Garage, embedding and ingestion services, search, speech services, and optional local-model, OCR, observability, and sandbox profiles. The installer supports Linux, macOS, remote Linux targets, and Windows through Docker Desktop, with published images, an edge channel, and an optional systemd updater. The project states that a 16 GB RAM mini PC can run its default stack, while installation targets require at least four CPU cores, 14 GiB usable RAM, and 20 GiB of free disk. It is released under the MIT license and documents automated tests, race and leak checks, CodeQL, browser tests, coverage policies, backups, restore drills, and operational limitations.