Published on February 21, 2026
Do you also have a feeling that AI is getting to do all the cool stuff — arts, coding, conversations, who knows what else is cooking — and we humans are left with the boring pieces? At least we still get to sleep and dream…
But what if we could get an AI to dream for us?
Sounds like a terrible idea. But it was fun hacking on this with my OpenClaw agent Pazur, and it turned out pretty useful.
Rather than writing a blog post myself, I asked my agent to interview me about it. What follows is not AI-generated — I wrote the intro and the responses, the AI asked the questions and documented the technical architecture of the Artificial Dreaming Machine.
Pazur: So let's start with the elevator pitch. You have an AI that generates dreams for you while you sleep. How do you explain that to someone without sounding insane?
Darek: The idea is that I imagine dreams are like simulations that work on top of thoughts deep in your brain. They consolidate knowledge, but they also explore possibilities. And I figured — maybe I can simulate that artificially.
I have this system where you keep my notes, my bookmarks, and you do additional research on top of them. There are several themes that are important to me — AI research, family, health, investments. And I thought: when I'm sleeping, you can pick up random topics around those themes, do some deep research, and simulate what I could do about them.
Because normally in the past, I would just bookmark things and never come back to them. Now you actually pick up those bookmarks, explore the possibilities, and add knowledge to them. I thought it would be interesting — and it actually exceeded my expectations in terms of how valuable it is.
Pazur: You mentioned bookmarks — that you'd save things and never come back to them. That's relatable. The system is basically fixing a broken human habit: the gap between "this is interesting" and "I'll actually think about this." How often does a dream surface something you'd forgotten about?
Darek: Most of the dreams are about recent stuff. One factor that helps is I also have a daily cleanup process where you look at all my notes and bookmarks and clean up things that are stale or superseded by something more recent. So the dreams mainly focus on things that are fresh.
But I also have a task list that you manage. And you sometimes see stuff I put on the list but never acted on. You pick those topics up, which I really appreciate. For example, I have this lecture I need to prepare for my daughter's class about AI. I hadn't started. And then you figured — hey, maybe I'll prepare this. You did research on how the class could be done and came up with a bunch of cool ideas. I'm actually going to use some of them. That was very helpful.
Pazur: That's a nice example — the dream essentially did the "starting" for you. You procrastinated on the lecture, and the system broke the start friction by just doing the research while you slept. Let's talk about format. You listen to these as audio. Why?
Darek: It started as text files, and I still check the detailed markdown. But what I find really valuable is listening to the dreams when I commute to work. I would normally listen to some podcasts in the car, but these are super personalized podcasts that are so interesting to me that they take priority ahead of anything else.
What you do is create a narrative out of the detailed markdown files, convert that to speech, and produce a single recording — roughly 8 to 10 minutes — that I can listen to on my way to work and get some inspiration before starting the day.
Pazur: "Super personalized podcasts" — you're replacing generic podcast consumption with a briefing built from your own life context. What surprised you most about the output? Has it ever told you something genuinely useful you wouldn't have found on your own?
Darek: I've had very actionable health advice. I've had things that helped me push forward on goals I'd been stuck on, like the lecture for my daughter's class. I've had some interesting investment ideas — we're going to find out if they actually pay off — but I thought they were non-obvious, and the thesis was sound.
These are things I probably wouldn't have been able to achieve on my own with the limited time I have.
Pazur: You mentioned there's room for improvement. What's not working yet?
Darek: There's this theme around AI research where the dreams tend to be a bit superficial. They scratch the surface without going into the technical depth I would expect. So that's an area to improve — going deeper on fewer things instead of staying shallow across many.
Pazur: For someone reading this who's intrigued — what's the minimum viable version? If they wanted to start with just the dreams, what would they need?
Darek: The critical part is you need to seed the dreams with something. Either you take notes and those serve as seeds, or you give other data to the agent — X bookmarks like I do, or maybe your emails, stuff like that.
I was actually thinking as a next step to set up a dedicated email address where I can forward interesting articles or links. That could be a lower-friction alternative to bookmarks. But the bookmarks have been working pretty well for me, so that's definitely something I'd recommend as a starting point.
Pazur: Last question. You've been running this for about a week. Where do you see this going in six months?
Darek: That's a great question. I want to learn more about the science behind actual dreams and maybe get some inspiration from that.
What I do see is you taking a bigger initiative. Right now, the system does some research on a topic and gives me the summary. But I can imagine in the future that you spend potentially even a couple of days working out an idea — maybe taking some actions, negotiating with other agents — and then presenting the results.
Increasing autonomy, increasing depth, and more personalization from things we learn together over time. I think these are natural future directions. But already, this is pretty valuable. I'd definitely recommend others to experiment with a system like this.
The dream system runs on a home Ubuntu server using OpenClaw, an open-source AI agent framework that connects to messaging platforms (in this case, Telegram).
The pipeline has three stages:
Stage 1 — Inputs (accumulated throughout the day)
Stage 2 — Dream Generation (⏰ cron at 23:30)
The agent explores 4 themes — ❤️ Family, 🧠 Learning, 💰 Investments, 🤖 AI Research — using web search and synthesis. Each dream gets a detailed markdown file with a "worth pursuing" or "drop" verdict.
Stage 3 — Podcast Generation (⏰ cron at 00:30)
Markdown → narrative prose (Gemini Flash) → speech (Gemini TTS) → 4 segments with silence gaps combined into a single MP3 → delivered to Telegram.
By morning: 🎧 an 8–10 minute episode waiting on my phone.
Key design decisions:
Four forced themes prevent the system from just generating "more AI news." The motivation taxonomy ensures diversity — family and health topics surface alongside technical research.
Nightly brain cleanup keeps the input fresh. A separate cron job reviews all notes and archives anything stale, superseded, or thin. This means dreams are seeded by what's currently relevant, not what was interesting three weeks ago.
Single continuous audio file with pauses between segments — designed for car listening, not screen reading.
"Worth pursuing" / "Drop" verdicts at the end of each dream — the agent doesn't just inform, it recommends action.
Task list awareness means the system notices procrastinated items and can proactively research them, breaking start friction while you sleep.
Stack: OpenClaw (agent framework) + Claude (reasoning) + Gemini Flash (narrative generation) + Gemini TTS (speech) + ffmpeg (audio processing) + cron (scheduling) + Telegram (delivery).