vellum-assistant

Vellum Assistant is an open-source personal AI assistant designed to be set up quickly and run continuously. Its central claim is a layered memory system: eight types (episodic, semantic, procedural, emotional, prospective, behavioral, narrative, shared), each with its own staleness window, hybrid dense plus sparse retrieval, and per-user and per-channel isolation. Structured items such as identity, preferences, projects and events are extracted from conversations with source attribution and deduplication, and embeddings run locally by default. Behavior is defined in a SOUL.md file; during onboarding the assistant observes communication style and writes its own personality files, keeps a per-user journal of reflections, and uses NOW.md as a scratchpad for current focus. A proactivity loop re-reads notes hourly, looks for unfinished or due items, and messages the user on the appropriate channel without interrupting active conversations. Security is built around actor identity (guardian, trusted, unknown) resolved once and enforced everywhere, with unknown actors unable to read memory, trigger tools or escalate. Credentials live in a separate process and are not exposed to the model, and tool calls run in a sandbox with a default-deny posture. Computer use is permission-gated: the assistant works in its own sandbox and, with approval, can read and edit files, run commands and drive a browser, with grants valid once, for ten minutes, or always. Channels include macOS, Windows, iOS, Web, Voice, Email, Telegram, Slack and Twilio, sharing one assistant and one memory. OAuth integrations cover Slack, Notion, Google, HubSpot, Linear, Discord, Twitter, Telegram and Twilio. Skills are plugins defined by SKILL.md and TOOLS.json that add tools and prompt sections at runtime, installable from a catalog, bundled, or dropped into the workspace. Deployment can be managed on the Vellum Platform or self-hosted from the same codebase. A CLI (installed via bun, or from source with setup.sh) offers commands such as hatch, wake, sleep, client, ps, terminal and upgrade. Model providers include Anthropic, OpenAI, Google Gemini, Fireworks, OpenRouter, MiniMax, Atlas Cloud and any OpenAI-compatible endpoint, with local models via Ollama. The project is MIT licensed and includes documentation on architecture, security, features, API and development workflow.