A few months ago, I switched back from Claude to Codex, and I've been on an OpenAI IV drip ever since. The best thing about this switch is that Codex doesn't have a five-hour limit, so you can just keep going. You almost never hit your weekly limits either, because Codex's lead developer regularly hands out "resets" when new things roll out. A bit like a drug dealer giving away free samples every now and then.
For most of my work, I now use GPT-5.6 Sol. For bigger questions, I sometimes ask Fable 5 to take a look as a second opinion. That combination, together with the Codex resets, means I've been using AI without limits for practically the entire summer. And as usual, I got carried away. Because what can you actually do with unlimited intelligence at this level?
My previous post was about handing things off to AI, and whether we dare to. Let's just say I've been testing that boundary. "On an OpenAI IV drip" is a pretty accurate description of what I've built.
The problem is context
Every task I take on, question I get, or bit of information I need now starts in Codex. Codex figures it out and comes back with the result. My role in this process is to make sure it has the right context to answer my request. That's also why I do everything in Codex. It keeps getting better the more you use it. A one-line bug fix? Codex. Change the color of a button? Codex. What day is it today? Codex.
Yes, this is overkill. I'm also very aware that somewhere, a data center is humming away for these useless requests. But hear me out.
First, I'm getting a very good sense of how these models work. Whenever Codex gets stuck on something, I immediately make sure it learns from that so the same thing doesn't happen again. Codex can also analyze your previous chats. That makes it easy to turn recurring tasks into Skills. If you use Codex from idea to deployed feature, you can use those analyses to adjust your setup so each new feature goes faster.
Give Codex enough context about your intent, your work, and the result you want, and it can usually handle the execution. My role then shifts from doing everything myself to setting direction, gathering context, and judging whether the result is right. In practice, everything really is a context problem.
The problem is that a lot of context still comes from outside Codex. An article I read, some quick feedback from a colleague, or a short call I forget to record. If I don't explicitly tell my agent about it, it doesn't know. So I started a new experiment to solve that. Its working title is "Remember".
Remember keeps track of everything I see, hear, and do on my Mac, and turns it into usable context for my agents. During the day, it builds a timeline of what I've been doing and captures conversations that used to get lost. The colleague who gives you some quick feedback, the bug that flew past in Slack while you were working on something else, or the few things you promised your boss that morning. Everything gets remembered and turned into context, so your agents understand you better.
In my own setup, I even keep audio recording on at all times, because I wanted to test what happens when truly nothing gets lost. Basically spyware.
It's a safety net for everything I do. AI will remember it for me anyway. With this much context, prompts like these are suddenly enough:
- Handle Walter's feedback.
- Give me an overview of what I did yesterday for stand-up.
- Fix yesterday's contract based on Tom's feedback.
Not all context is equally important. Remember can collect the explicit commitments I make in my Inbox. From the timeline, I can select a task and ask my agent to turn it into a Skill. Knowledge that might be useful later ends up in my personal wiki, a kind of long-term memory for your AI. I'm also experimenting with a shared wiki that lets you share information with colleagues, for example. That way, different agents working on the same project can draw on the same data source.
"Is it on?"
The reactions I get are interesting. When I explain the vision of a second brain and an AI that no longer needs me to supply context, people see a ton of potential. Then they see the microphone icon is always active on my Mac, and they get nervous. Even after I explain that the audio stays local, or say I'm experimenting with local models to run everything offline, it still worries them. Offline says nothing about whether the other person consents to being recorded.
I find the privacy side of the app fascinating. We store everything with Google, record conversations in Teams, connect agents to email and Slack, and paste all kinds of sensitive information into chat windows to write emails faster. And all of it goes to cloud servers, no less. So why does that little microphone icon suddenly feel different?
I still don't quite know what I want to do with this. On my own Mac, it feels like a fun experiment. But if I worked at a company and my boss rolled this out, I wouldn't be so chill about it. Maybe everything stays local for now. But who can guarantee that I won't have to explain my own work timeline to HR in two years?
At the same time, this idea doesn't go away if I stop building it. I see all kinds of signals that tell me this is where we're headed. OpenAI took a step in this direction with Computer History too. It's still opt-in for now. I think that's mainly because this works so well as an intermediate step. You let AI figure out for itself what it can take over for you. It solves a real problem and makes the switch to working with AI much easier.
At the very least, this experiment proves that hooking yourself straight up to an OpenAI IV drip opens up a ton of new possibilities. Giving AI more access to your life demonstrably leads to better work. But is it a second brain, or just spyware in disguise?
If you want to experiment with it, you can check out Remember here. Data stays local by default, and you can use its privacy controls to turn things off. I don't make any money from it and have no interest in your data. You can read here what the app stores.