Vellum Assistant is an MIT-licensed, TypeScript personal AI assistant for individuals and small teams who want a persistent, self-improving agent that runs across desktop, mobile, web, voice, email, Telegram, Slack and Twilio without building the memory and orchestration layer themselves.

What it is

Vellum Assistant is a personal AI assistant that keeps one memory and one identity across every channel it is connected to. Behaviour lives in a file called SOUL.md, which the assistant writes during onboarding after observing how the user communicates; a per-user journal of reflections and a scratchpad named NOW.md hold current focus and active threads. Memory is split into eight distinct types — episodic, semantic, procedural, emotional, prospective, behavioural, narrative and shared — each with its own staleness window, hybrid dense and sparse retrieval, and per-user and per-channel isolation. Embeddings run locally by default, and structured items such as identity, preferences, projects and events are extracted from conversations with source attribution and deduplication. Channels include macOS, Windows, iOS, Web, Voice, Email, Telegram, Slack and Twilio.

The concrete problem is setup cost. Anyone who has stood up a personal agent on OpenClaw, Hermes Agent or Claude Code knows how long the configuration takes and how often the agent has to be rebuilt before it behaves correctly. Vellum aims to deliver the working result out of the box from a single download, replacing the hand-maintained combination of a SQLite database and a Markdown file that self-built assistants usually end up as. It also replaces hand-rolled OAuth token refresh for Slack, Notion, Google, HubSpot, Linear, Discord, Twitter, Telegram and Twilio.

Key capabilities

• Eight memory types with per-type staleness windows, hybrid dense and sparse retrieval, and per-user and per-channel isolation.
• Local embedding generation by default, so conversation content is not required to leave the machine for retrieval.
• Actor identity resolved once as guardian, trusted or unknown and enforced everywhere; unknown actors cannot read memory, trigger tools or escalate.
• Credentials held in a separate process and never passed to the model, with every tool call executed in a sandbox and denial as the default.
• Proactive hourly review: the assistant re-reads its notes, finds unfinished or due items, and messages the right channel without interrupting an active conversation.
• OAuth integrations for Slack, Notion, Google, HubSpot, Linear, Discord, Twitter, Telegram and Twilio.
• CLI commands including vellum hatch, vellum wake, vellum sleep, vellum client, vellum ps, vellum terminal and vellum upgrade, with an assistant ID accepted as a second argument when multiple assistants exist.

Who uses it and how

• Individuals who want a single assistant reachable from macOS, Windows, iOS, Web, Voice, Email, Telegram, Slack and Twilio, sharing one memory across all of them.
• Terminal-oriented users and contributors who install with bun install -g vellum, run vellum hatch, and drive the assistant through vellum client or vellum terminal.
• Self-hosters who want everything running on their own machine, including local embeddings and a sandboxed tool runtime, rather than a hosted service.
• Teams that need per-user and per-channel memory isolation plus actor-based access control, so untrusted actors cannot read memory or trigger tools.
• Users who want to reach a self-hosted assistant from a phone or another computer by opening a tunnel and pairing the device.

Getting started

Sign up at vellum.ai or download the desktop app for macOS and Windows, then hatch an assistant; the CLI path is bun install -g vellum followed by vellum hatch. Managed mode signs in through Vellum Cloud with no local runtime, while local mode runs everything on the machine.

How it compares

Against OpenClaw, Hermes Agent and Claude Code, the difference the README draws is time to a working assistant: Vellum ships eight memory types, OAuth integrations and channel routing as configuration rather than as work the user assembles. It is MIT-licensed and can run fully self-hosted on the same codebase and data model as the managed option.

When to use it — and when not to

A self-hoster takes on the local runtime, local embedding generation and the sandboxed tool execution environment, so this suits someone comfortable operating a service on their own machine. Do not pick it if a stable Windows desktop build is required: the Windows download is currently a dev build connected to the development environment, with x64 and ARM64 installers. The project also carries a notable open-issue count of 179 against 1280 stars and 186 forks, and the README excerpt supplied here is partially truncated in its infrastructure and security section, so some deployment detail must be read from the linked documentation.

bun install -g vellum
vellum hatch

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

vellum wake        # start services
vellum sleep       # stop services, keep data
vellum client      # interact through the terminal
vellum ps          # view running assistants
vellum terminal    # open a shell into a managed assistant container
vellum upgrade     # upgrade to latest version