vellum-ai/vellum-assistant

What it solves

It eliminates the complex setup and frequent resets required to create a functional personal AI assistant. It provides an out-of-the-box solution that learns user behavior, remembers critical information across multiple platforms, and takes proactive actions without requiring manual maintenance of memory files.

How it works

The assistant uses a sophisticated memory system comprising eight distinct types (episodic, semantic, procedural, emotional, prospective, behavioral, narrative, and shared) with hybrid dense and sparse retrieval. It manages identity through personality files (SOUL.md) and a focus scratchpad (NOW.md). For proactivity, it periodically reviews its notes to notify users of unfinished tasks. It operates across various channels (macOS, iOS, Web, Voice, Email, Telegram, Slack, Twilio) and integrates with numerous apps via OAuth. Security is handled through a sandboxed environment where tool calls and credentials are isolated from the model.

Who it’s for

Users seeking a personal AI assistant that can be self-hosted or managed, and those who want an AI that evolves with their specific work habits and communication style across multiple devices and apps.

Highlights

• Multi-layered Memory: Eight specialized memory types with per-user and per-channel isolation.
• Cross-Channel Continuity: A single assistant and memory shared across platforms like Slack, Telegram, and macOS.
• Proactive Engagement: Hourly reviews of notes to alert users to pending tasks.
• Sandboxed Execution: Permission-gated computer use (file editing, browser driving) and isolated tool execution.
• Broad Provider Support: Compatible with Anthropic, OpenAI, Google Gemini, and local models via Ollama.