# Vellum
Vellum (vellum.ai) is a personal AI assistant; it calls itself "your Personal Intelligence." One assistant, one memory, reachable from a macOS app, iOS, the web, a CLI, Telegram, and Slack. It learns how you work and acts before you ask.
The pitch that sets it apart from a chatbot: the assistant has its own identity. Its own name, its own email, its own accounts and logins. Its behavior lives in a `SOUL.md` file, and during onboarding it watches how you communicate and writes its own personality files. It keeps a journal of reflections on past interactions.
Memory is structured, not a flat chat log. Items like identity, preferences, projects, and events are extracted with source attribution and deduplication, then retrieved with a hybrid of semantic and lexical search.
It ships with 60+ skills and can be taught more. Autonomy is configurable, from "ask before every action" to fully autonomous. Credentials stay in the macOS Keychain (self-hosted) or an isolated vault (managed); the model never sees them, and Vellum doesn't train on user data.
The roadmap points further in the same direction: subagents for parallel work, Teleport (handing a live session between local and cloud without losing memory), Google Meet and iMessage integrations, a paid Pro tier, self-improvement without fine-tuning.
Vellum raised $25M. Backers include Dharmesh Shah, Arash Ferdowsi, Rebel Fund, and [[Y Combinator]]. The assistant core is open source at github.com/vellum-ai/vellum-assistant.
**Why it matters here:** this is the OSK agent system, productized and hosted. SOUL files, accumulated memory, a skill library, subagents, journals; the parallel is almost exact. Worth watching as both a competitor and a reference design.
## References
- https://www.vellum.ai
- https://www.vellum.ai/roadmap
- https://github.com/vellum-ai/vellum-assistant
## Related
- [[AI Assistants]]
- [[AI Agents]]
- [[AI Subagents]]
- [[Memory]]
- [[Large Language Models (LLMs)]]
- [[Context Engineering]]