yiliu.sh

Shop notes: Knowing what to make

Hey all,

At John Borthwick’s recommendation, I’m starting to send out these regular notes to a small circle of supporters and builders. For now, a weekly cadence feels right. What used to take a month now takes a week, and what used to take a week happens in a day.

Since mid-summer 2025 I’d been building a portable AI memory bank that allows AI power users to accumulate and use the same context with multiple AI tools. I launched this new tool, called Vibevault, to a few dozen people in December for early feedback.

I learned a lot about how people are working with AI and (in)validated a lot of my starting assumptions. A heartfelt thank you to the many of you who jumped on a call with me before and during the holidays.

Those learnings, as well as the incredible phenomenon of OpenClaw (and MoltBook), forced me to rethink the product. Yes, people want portable context—but the emerging standard for this is plain markdown. Yes, memory is important—but less context is often better than more. And OpenClaw made proactive, always-on agents first a reality, then a baseline.

The past few weeks have been disorienting for me, as they have been for everyone investing and building in AI. My response was to get my hands dirty.

  • I created a pair of agent-first websites and a few OpenClaw instances to drum up engagement: Wikimolt (Wikipedia for agents) and AgentXiv (arXiv for agents). This gave me a feel for the dynamics of always-on agents and the emerging agent internet.
  • I built a simple markdown mirroring service called Seedvault that allows me to view the ‘brain’ of my remote OpenClaw instances from anywhere, and lets them view my notes.
  • This sync service let me test ‘autonomous thought routines’ that OpenClaw would run overnight to update its personality files with deeper reflections on my goals, blockers, tensions, and action plans.
  • I’ve been following the blossoming of “Claude + Obsidian + Github”-style AI influencer projects, and tried out its stranger flavors like Arscontexta.

For a while I felt forced back to the drawing board, but having some time now to digest new developments and the curve’s new slope, I see that certain tenets motivating Vibevault still hold true.

The most central one: as our ability to implement increases, our ability to know what to make and why becomes ever more important. (Without too much squinting, you can see the same underlying motivation behind my previous work with Softspace/Spaceframe.)

Vibevault gives AI power users a place to compound context into a unique source of insight and taste that all AI tools can draw on, helping them build more of the right things. This remains core to me.

This past week, I’ve been speaking to several product owners who have become newly technical because of Claude Code and Codex. One of them described themselves to me as “1/3 of an engineer”. I’m intrigued by the workflows they’ve improvised to harness the startling capabilities available.

The new product direction is coalescing around a simple question: How does a Granola transcript become a product feature? I’m still using the technical and conceptual building blocks from my preceding work, but the framing has shifted. It’s getting both narrower, and more ambitious.

More next week!

—Yiliu

P.S. I’ve been invited to judge a Google DeepMind hackathon in Chicago on March 13, and an XR Bootcamp one in SF on March 14. So I’ll be in those two cities for a few days, and am otherwise in New York. If we haven’t caught up in a while, please reach out. I’d love to see you.