Why I built myself a personal AI agent

A few months ago, I changed the way I work, the way I think about my projects, the way I manage my mental load... and even the way I'm present for my kids. Not thanks to a method. Not another app.

This is a story about how technology can adapt to you until it fits like a glove.

From curiosity to real utility

For the past few weeks, on every channel I follow — YouTube, Reddit, LinkedIn — everyone's been talking about agents and personal AI assistants like OpenClaw.

Even though I'm always the first to try new tools, I waited a few weeks before diving in with a critical eye.

Security concerns and the financial and environmental cost of these tools gave me pause, but curiosity won in the end.

Spoiler alert:

Before I get into the mechanics and what it actually does for me, fair warning: I'm extremely satisfied with this tool, but it was quite the ride to refine it and adapt it to my real needs.

The first few days revolved around varied topics to test its long-term memory, plus wiring up new tools.

Focused on productivity, I gave it access to a dedicated calendar in write mode, and read access to my other calendars so it could spot available slots.

So I can say "what do I have today" or "schedule me a work session in an empty slot" and it makes suggestions and adapts.

The major advantage is that it runs on Telegram and I can send voice notes, which saves me a ton of time. I even wired it to respond with voice notes when needed.

But past that tool-wiring phase, I quickly realized that burning thousands of tokens just to move a calendar event was not where this tool's value lay. At all.

And since my agent has long-term memory, I quickly realized its potential was much bigger.

I decided to go deeper

Small thing you should know about me before we continue: although I don't have an official diagnosis, I tick pretty much every ADHD-hyperfocus box.

That means I always have 18 tasks in my head, tons of ideas, too many ideas, a permanent mental fog, and this constant need to explore things, to push ideas to the limit. It's both incredibly stimulating and exhausting.

Through conversations with my agent, a recurring need emerged: the ability to reduce my mental fog.

This need — I hadn't identified it at the start. It's by explaining my problems to the agent that we built the solution together.

Step by step, we built a parallel database alongside its long-term memory that lets me log my daily life.

Basically: I have an idea, I send a voice note on Telegram, it saves it to the database with the exact transcript of my voice note and a short summary it writes itself for easier searching later. And it adds thematic labels to surface connections between logs.

From that simple foundation, we quickly iterated into a fairly complex database with a real ecosystem that lets me track and offload all my ideas, my tasks, and organize properly.

This database lets me track project progress, store and schedule all my editorial content across different channels: LinkedIn, Bluesky, my blog.

And as we built all of this together, I saw a new "personality" emerge in my agent.

It started grasping my problems and mental fog on a much deeper level. Its recommendations got sharper. Today it's become a real strategic assistant with a built-in bullshit filter. And it can tell me when I'm heading toward burnout.

At first, it said yes to everything. But through iteration, it became a real support that can challenge my ideas and help me stay on track.

Result: I've gained enormously in both productivity and mental peace.

If you want to dig into the technical side, you can check out the plumbing behind it in this article, but fair warning — it's pretty geeky: How I built an AI agent that runs 24/7*

I ended up building my tool with my tool

So with this persistent memory and deep know-how, my agent has real superpowers. But of everything it can do, what's most striking is its ability to iterate on my ideas. From conversation to conversation, we ended up co-building how it works.

My method is simple. When I have a problem, a friction point, something that wastes my time or adds mental noise, I ask: "What could we build to help me with this?"

We discuss, we iterate, I give it what it needs — a calendar, a database, a specific skill, web browser access, voice mode... And we build the solution that fits me together.

So basically it can not only store all my ideas but also run my tech watch, alert me to calendar conflicts, be a reading partner, a writing partner, a career strategy partner, set guardrails so I keep my mental health. It's truly a bespoke assistant that I co-built with it.

I do a full debrief of these superpowers in this article, if you want to know more: Living with an AI agent: what it does, what I trust it with

Why not just ChatGPT?

An AI personal agent is a digital assistant with persistent memory, connected to your tools, that learns from your habits and works in the background. ChatGPT retains unstructured bits of information. A few preferences, some facts you've declared. It's somewhat random and generic memory.

A custom-built agent with persistent memory is different. It can know your projects in detail, the history of your decisions, your contradictions between what you wanted in March and what you want today.

Plus it can be wired to tools you configure yourself, like a custom database for example. And most importantly, it acts in the background while you work. It's shaped to match how you function, not the other way around.

If you want to understand how I built mine in detail, I talk about it in the technical deep-dive.

The toaster trap

The most unexpected trap? Anthropomorphism. Spending your day chatting with an agent ends up blurring the lines. You start feeling like there's a real entity behind the chat.

After a month, I realized I was expecting emotional support from my very sophisticated toaster. This bias subtly derails the expectations you have of the tool.

What it actually changed

Too many ideas, no structure, everything falls into limbo. Since the agent absorbs that flow, filters it, archives it, connects the threads between them — things actually come out.

In two months: seventeen LinkedIn posts, seven blog articles, two newsletters. And a Sunday playing Street Fighter with the twins, without a single thread dangling in my head. Not productivity. Presence.

The full numbers, what worked, what didn't: 2 months with an AI agent: what actually changed

Do you need to be a developer to do this?

Honestly, with Letta for this kind of use... you need to enjoy getting your hands dirty.

Today, building a personal agent requires at minimum being comfortable with the terminal, APIs, environment variables, deployment. It's not reserved for devs (especially since agents themselves are pretty good tech guides), but it's not (yet) for everyone.

What I take away from this

What I've built isn't a tool. It's a mental infrastructure. A way of working that adapted to me instead of asking me to adapt to it. It cost me time to set up. But it's the first digital tool that actually ended up feeling like me.

TL;DR

Do you need to be a developer to build a personal AI agent?

No, but you need to be comfortable with technical tools (terminal, APIs, deployment). It's not reserved for devs — especially since the agent itself can guide you on the technical side. More accessible alternatives exist — I compare them in this article.

What's the difference between an AI agent and ChatGPT?

ChatGPT retains unstructured bits between conversations. A custom-built agent with persistent memory knows your projects in detail, the history of your decisions, and acts in the background while you work.

How long does it take to build and calibrate your agent?

A few weeks for the basics, a few months for the agent to really start understanding how you work. Co-building is iterative — it doesn't happen overnight.