A personal AI assistant designed to evolve with the user, Vellum Assistant learns work habits, remembers important information, and acts proactively across macOS, Telegram, and Slack. It offers a personalized and extensible AI experience, aiming to enhance productivity by anticipating user needs and managing information intelligently. The target audience includes power users and developers seeking a sophisticated, adaptable AI companion.

How It Works

Vellum Assistant employs a hybrid approach to memory management, utilizing structured items like identity, preferences, and projects, extracted with source attribution and deduplication. It features a hybrid retrieval system (dense + sparse) for semantic and lexical ranking, with configurable staleness windows and per-user/channel isolation. Embeddings run locally by default. The assistant develops a unique identity through its SOUL.md behavior file, learns communication styles during onboarding, and maintains a per-user journal and ephemeral scratchpad (NOW.md). Proactivity is driven by hourly self-checks that identify unfinished tasks or upcoming deadlines, triggering relevant messages without explicit user prompts. Security is paramount, with a fail-closed trust engine resolving actor identities (guardian, trusted, unknown) and enforcing strict permissions, ensuring untrusted actors cannot access memory or trigger tools. Credentials are managed in a separate process, and all tools operate within sandboxed environments.

Quick Start & Requirements

Installation via CLI:

bun install -g vellum
vellum hatch

Installation from source:

git clone https://github.com/vellum-ai/vellum-assistant.git
cd vellum-assistant
./setup.sh
vellum hatch

The desktop application is the primary focus, with the CLI available for advanced users and non-macOS environments. Prerequisites include bun for CLI installation and git for source installation.

Highlighted Details

• Evolving personal AI assistant with structured memory and learned identity.
• Proactive reach-outs based on learned context and tasks.
• Robust security model with fail-closed trust engine and sandboxed tools.
• Multi-channel support (macOS, Telegram, Slack) with shared memory.
• Flexible model support (Anthropic, OpenAI, Gemini, Ollama) and local embeddings.
• Extensible via sandboxed, manifest-driven skills.

Maintenance & Community

Vellum Assistant is an open-source project by Vellum AI, a for-profit company that also offers a managed Vellum Platform. Contributions are welcomed, with a contributing guide and Code of Conduct available. Community support is accessible via Discord.

Licensing & Compatibility

The project is licensed under the MIT license, permitting free use, modification, and distribution, including for commercial purposes.

Limitations & Caveats

The project's primary development focus is the macOS desktop application. While a CLI is provided, it is designated for advanced users and non-macOS environments, suggesting potential limitations or a less polished experience on other platforms.