Why Petals exists

In 2015 I wrote a short manifesto about technology and attention. The thesis was simple enough to fit on a poster: focus is power, and the major tech companies generate revenue by destroying yours. Every copy-paste is a chance to eliminate time you spend being a machine. Let the machines be machines and let us be humans.

I believed it then. I still believe it. What I didn't anticipate was how much more pointed the problem would become once the machines got good at conversation. Strange, how that goes — the better the tools got at talking, the sharper the old problem became.

---

The AI tools that emerged in the early 2020s were remarkable. I used them constantly. But I kept noticing a pattern: the more I leaned on them, the more time I spent re-explaining myself. Every new session began with context-setting. Who I am, what I'm working on, what I care about, what I don't. The tool was capable, but it was stateless — a brilliant stranger I had to re-introduce myself to, over and over. You know the feeling. You've had this conversation before, many times, with the same capable stranger who never remembers having it.

This is what I started calling cognitive debt. Not in a metaphorical sense. I mean it literally: every time a tool forgets what it already knows about you, you pay an interest installment. You spend five minutes at the start of every conversation reconstructing a context that should already exist. Multiply that across a year, across every tool you use, and the cost is enormous — not just in time but in the quality of the thinking that emerges from those conversations. A tool that doesn't know you can't ask the right questions. It can only answer the ones you already know to ask.

The obvious fix seems to be: make the AI remember. And the major products all added memory. ChatGPT, Claude, Gemini — they all have some version of it now. But I kept finding the implementations unsatisfying in a specific way.

The memory was invisible.

You couldn't see what the model had retained. You couldn't edit it. You couldn't correct a misremembering or delete something that no longer held. The model's picture of you existed somewhere, shaping its responses, but the picture was hidden from you. ChatGPT's own documentation says it "does not necessarily include everything ChatGPT may remember." That's a polite way of saying: we know things about you that we're not showing you.

An assistant with a hidden picture of you isn't a tool. It's something closer to a surveillance feed that also answers questions.

---

I want to be precise about what I find objectionable, because it's not memory itself. Memory is exactly right. The problem is memory without transparency, without control, without the ability to inspect and correct what the tool thinks it knows.

The 2015 manifesto was about focus — about not letting machines control us, about making them work for us. The principle maps directly onto this. A tool that holds a hidden model of you has, in some small but real way, inverted the relationship. You're not using it; it's using you.

Petals is my answer to that inversion.

The memory in Petals is a graph you can see. Every claim the system has extracted about your life — a person, a project, a recurring concern, a stated value — shows up as a node. You can read it. You can edit it. You can delete it. When two memories contradict each other, the system surfaces the conflict rather than silently picking one. When you want to understand why the system believes something about you, there's a trail back to the source: the conversation, the document, the note where that belief came from.

That transparency isn't a feature bolted on afterward. It's the architecture. I built it that way because I don't think an AI assistant is trustworthy unless it shows you what it knows.

---

Memory is the foundation, but it's not sufficient. An assistant that knows you but can't act on your behalf is a very smart notebook. What I wanted was something that could reach into the tools I actually use — my calendar, my email, the services I depend on daily — and do things there. Not autonomously, not without permission, but with the kind of targeted agency you'd want from a capable colleague.

The design principle I kept returning to was consent at every threshold. The assistant acts as far as you allow, and no further. You define the scope; you approve the gates. The model doesn't reach into your calendar because it decided to — it reaches there because you connected it and said it could.

This is the difference between a tool that works for you and a product that works on you. The big AI products optimize for engagement. More sessions, longer sessions, deeper entrenchment. That's the business model. Petals optimizes for the opposite: calm, purposeful use that reduces the friction in your day without manufacturing new reasons to open the app.

---

I'm aware this sounds like a positioning statement. So let me try to be more honest about what I mean.

The conscious technologist's manifesto was written in reaction to a specific moment — the peak of the attention economy, the smartphone notification spiral, the first wave of infinite-scroll feeds. The enemy was distraction-as-product. The response was ruthless elimination: cut what doesn't matter, invest everything in what does.

The enemy in 2026 is more subtle. It's not distraction in the crude notification sense. It's dependency. It's the quiet erosion of your own context, your own memory, your own understanding of where things stand — because you've offloaded all of it to a tool you can't fully inspect. It's cognitive debt compounding silently while you think you're being productive.

The assistant I wanted was one that could hold context without holding it hostage. That could remember for me without thinking for me. That could act on my behalf without acting without my knowledge.

I built Petals because that assistant didn't exist yet.

---

The soul line I keep coming back to is this: it remembers so you can think. It never thinks so you can forget.

The first part is the product. The second part is the constraint. An assistant should reduce the cost of retrieving what you already know — not replace the process of thinking through what you don't.

Wu wei is a Taoist concept that translates roughly as "acting without forcing." The ideal action is the one that meets the situation without adding unnecessary weight to it. I think about that a lot when I'm deciding what Petals should and shouldn't do. The best intervention is usually the quietest one. The right memory surfaces without announcing itself. The right action happens without requiring a conversation about whether it should.

A calm tool that knows your life because you chose to share it. That acts only as far as you allow. That asks nothing more than the moment requires.

That's what I'm building. That's why.