AI Trends

This VC built an AI chief of staff on OpenClaw — and never wants to go back

Remy Gieling

April 7, 2026

min read

This VC built an AI chief of staff on OpenClaw — and never wants to go back

An investor let AI manage his email, calendar, and relationships — and never wants to go back to a human assistant.

Ryan Sarver, ex-Twitter and Redpoint Ventures, shares his complete setup for “Stella”: an AI assistant that triages email, prepares meetings, manages fundraising with 100+ contacts, and improves himself every week. His claim: better than any human chief of staff he has ever hired.

Ryan Sarver isn't the first tech entrepreneur to write enthusiastically about AI assistants. But his extensive post on X stands out for its depth. Sarver — who built the developer ecosystem as Director of Platform at Twitter and was subsequently a partner at Redpoint Ventures — now runs his own Kelp fund. In the middle of fundraising, sitting on boards and active as an angel investor, he built an AI chief of staff that he calls “Stella” on the open-source platform OpenClaw.

Not a toy, but a fully operational system. And it starts with a fundamental design principle.

Memory is the core, not the model

Most people who work with AI assistants rely on call history as memory. Sarver calls that a recipe for frustration: session memory disappears, fills up, or misses crucial context at exactly the wrong time.

His solution consists of two layers. The first is a daily log file — one markdown file per day that automatically tracks what happened that day: meetings, decisions, tasks, context from conversations. A script retrieves this from sessions and writes it away without manual intervention.

The second layer is a central file (Memory.md) that Stella manages and updates herself. Every processed meeting, every trickled email, and every tracked task continuously feeds this long-term memory. Without this layer, you have a capable amnesiac assistant, Sarver writes. With this layer, you have something that looks like a colleague who has been working alongside you for months and never forgets anything.

A conscious choice of architecture: everything lives in flat markdown files, not in a database. Sarver can open, read, and correct any file. Everything can be backed up to Git. There is no layer of abstraction between him and what his world's assistant understands — and that's what makes him trust the system.

What Stella does every day

Stella combines tasks that you would normally distribute across multiple tools and people.

In the area of email and calendar she scans multiple Gmail accounts, filters what action requires, and drops the rest. In addition, she automatically retrieves expense receipts for quarterly reports, generates travel schedules from booking confirmations, and prepares follow-up emails in Sarver's own writing style.

Twice a day, Stella sends a briefing via WhatsApp: at 9 a.m., a morning letter with top priorities, overdue tasks and the daily agenda, and at 6 p.m., a summary of what happened that day, what has stalled and what needs attention tomorrow.

The most impressive application is the fundraising pipeline. Sarver manages relationships with over a hundred LP contacts in multiple countries. Stella keeps track of the entire pipeline, knows where each relationship stands, and prepares a letter for each meeting: she researches the fund, scans recent publications by the partners, makes connections with Sarver's investment thesis, and provides customized discussion tools. For ongoing relationships, she knows exactly what has been discussed, what has been promised, and where the sensitivities lie.

Kaizen: the system that improves itself

The part that distinguishes Sarver's setup from most OpenClaw builds is the weekly improvement cycle — in his words, “kaizen” for AI.

Every Friday, a cron job does an investigative task. Stella scans the OpenClaw community, looks for new patterns, and sees what other builders are doing. On Sunday morning, Sarver and Stella will discuss the findings together: what are the best ideas, and what will actually be modified?

But the real strength lies in internal learning ability. If Sarver continues to correct something, or if a function provides more friction than value, that is recorded in memory and eventually comes up as an improvement proposal. Too much noise in a triage filter? Briefing format that doesn't land? Stella notices and suggests adjustments.

This is something a human chief of staff can't do at scale, Sarver writes. A person learns from working with you, but can't simultaneously scan what hundreds of other builders do and cross-compare that with your system every week.

The tool stack under the hood

For those who want to build this or understand what is technically required — here's what's under Sarver's setup:

Platform: OpenClaw (open-source, self-hosted on own hardware or VPS). Communication: WhatsApp as a primary channel for briefings and interaction. E-mail: Gmail integration (multiple accounts). Memory: Flat markdown files — daily logs plus a central Memory.md backed up via Git. Automation: Cron jobs for daily briefings, weekly research scans and continuous memory building. File format: Everything in markdown — deliberately no database, no proprietary storage.

An important note: this is not a plug and play solution. Sarver is a tech-savvy builder with years of experience in platform development. Setting up and fine-tuning this system requires you to be comfortable with the command line, markdown, cron jobs, Git, and iteratively adjusting AI behavior. If you're not, you'll probably need a pre-built OpenClaw persona like Atlas, or a similar managed solution.

What we think about this

Sarver's post is impressive as a blueprint. It shows what's possible when you treat AI not as a chat window but as an operational system — with persistent memory, proactive triage, and a machine learning improvement cycle. The architectural principles he describes (memory as a foundation, markdown as a transparent storage layer, continuous improvement) are universally applicable and are not limited to venture capitalists.

At the same time, nuance fits here. Sarver presents his setup as superior to human chiefs of staff, but compares apples to pears. A human chief of staff brings political insight, emotional intelligence, and judgment to sensitive situations — qualities that an AI system does not have. Stella excels in everything that is systematizable and repeatable: triage, briefings, pipeline management, pattern recognition. The real strength is probably not in replacement, but in combination.

For the Dutch market, this pattern is particularly relevant. More and more organizations are not approaching AI as a standalone tool, but as a layer that runs through business processes — AI as an operating system, instead of AI as a gadget. The shift from “I'm asking ChatGPT a question” to “an AI colleague with memory, initiative, and an improvement cycle” — that's exactly the move Sarver is demonstrating here. And it's the direction that any knowledge worker who wants to get serious about AI should take into account.

Sarver offers to make its entire system open source available if there is enough interest. We hope he does. Not because everyone has to literally recreate this, but because it raises the bar for what we can expect from AI assistants — and what happens if you take the time to really set them up properly.