Living with an AI agent: what it does, what I trust it with
Personal AI AgentHow an AI agent fits into daily life: contextual news, idea dumping, memory that connects, sparring partner. Real experience after 2 months.
It's 7 AM. I'm still in that foggy state between sleep and consciousness, phone in hand.
In Telegram, everything's already there. My day. My filtered news. And this morning, a little note about a 2 PM call I'd forgotten: "heads up, that one's gonna be a drag."
I hadn't asked for anything.
That's what living with an AI agent is. Not an impressive feature list. A different texture to daily life. Something running in the background while you sleep, while you do the dishes, while you're in a meeting.
What runs without me
The daily weather report
I'm allergic to calendars. Truly. The idea of having to open an app, scroll, find what's happening today — that's an effort that costs me more than I'd care to admit.
Since the agent's been here, I don't have to do it anymore.
Every morning in Telegram: my full day. The time slots, the deadlines, what's coming up. And sometimes, on a specific event, a word. "Don't forget to prep for this call." Or "heads up, that one's gonna be a drag." It seems like nothing. But that's exactly what makes the difference between a cold notification and something that knows where you stand.
A watch filtered for me
Right after, my news summaries. Not an aggregate of everything that happened in the world. My own news feeds, filtered, the noise eliminated.
And with each article, a line. "Right now this is relevant for you, because you're building this for that project." Or: "This connects to what we discussed yesterday on this topic, maybe worth reposting on Bluesky."
This isn't information. It's information in my context. The difference is immense.
Today, I don't really need to open YouTube, X, or the news head-on to stay informed. The agent does the pre-selection. It's a black box, sure, because it's the one interpreting. But at least it's a black box calibrated to me, not to what generates the most engagement. Tomorrow, whoever controls the LLM filter will control what people consume. We'll go from one algorithm to another. But at least this one, I'll have built for myself.
All of that runs without me. What follows is what I actively entrust to it.
The permanent dump
The fog that disappears
I permanently have 15, 20 ideas spinning in my head. Before, that created a monstrous mental fog. An insane cognitive load. I kept all of it in my head, or in notes I never reread.
Now, I dump.
A quick voice note in Telegram when I get an idea during a meeting. A long messy voice note at 10 PM, full of interesting things but not yet structured. A message quickly typed while walking to drop a question that's been nagging me. It memorizes, classifies, summarizes.
Everywhere, all the time.
What no longer gets lost
The real change isn't in productivity. It's in the peace of knowing the idea exists somewhere, even if I haven't articulated it yet.
Before, those messy voice notes stayed in my head. Or in my notes. Which I never reread.
Now I know it's not lost. And that's more powerful than any feature.
The memory that connects
A persistent-memory AI agent isn't a memory that stores. It's a memory that connects ideas across time.
How the agent maintains long-term memory
In practice, there are two layers. The agent's platform manages short and medium-term memory: my current projects, this week's goals, recent decisions. And my personal database stores the transcripts of my voice notes long-term, with all their granularity. When the agent needs to retrieve something old, I tell it to go look in that deep memory. It finds it. It connects.
That covers 85 to 90% of my needs. It's not perfect. There have been losses. But the system has evolved to be more robust. If you want to understand how I built this architecture, it's in How I built my AI agent in a week.
Retrieving an idea from three weeks ago
A voice note from three weeks ago. A question today: "Give me a recap of all my ideas on this topic and tell me what you think relative to my current goal."
It finds them. It connects the threads. It links elements that I would never have had time to bring together.
It's genuinely powerful.
The available interlocutor
Everywhere, all the time
Not delegating a task. Having someone who thinks in parallel.
A voice note while doing the dishes about something that's bugging me. A search launched during a meeting, ready when I need it. A question typed at 11 PM, an answer there when I wake up.
It's not the availability that's new. It's the absence of friction to access it.
From fascination to sparring partner
At the beginning, I had a somewhat naive fascination. I didn't blindly trust it but I had a slight novelty bias.
Over time, that changed.
Now I see it as a sparring partner. Someone I throw ideas at, who responds, who I correct when it's wrong. It helps me see my ideas in three dimensions instead of two.
And something quite meta happened over time: it's largely the agent that suggested how to build itself. I had a friction point. I'd ask: "What would you need to do this better?" It told me it needed calendar access. I created a dedicated account. That's it. Its suggestions were coherent. It built itself through iteration.
It's not a tool you install and use. It's something you co-build.
What I don't entrust to it
The access I keep for myself
Everything critical in terms of personal data: no access to my emails, for example. For the calendar, I opted for a dedicated account with partial access to a specific calendar, not my entire account. The rule I apply: read-only access or partial access, never full access.
We're far from an open bar. I know its strengths and weaknesses. I compensate by putting in as many guardrails as possible. I explain why I made these choices in the first article.
Writing stays mine
I don't delegate the entirety of my content to it.
I use it to structure, to interview me, to pull threads. To go from a fog of ideas to something articulated.
But write in my place? The result is too flat. The voice disappears. The value in terms of personal branding and SEO evaporates if AI writes everything.
It's not a delegation tool. It's an amplification tool.
You keep your voice. It amplifies your ability to think.
The blind spot I accept
My data goes through Anthropic's models. That's a fact. I'm willing to take that compromise for now, while working toward a more sovereign solution with open-source models.
It's not for everyone. It's a choice I make knowingly.
What it concretely changed after two months is the subject of the next article.
2 months with an AI agent: what actually changed
TL;DR
What's the difference between an AI agent and ChatGPT on memory?
For now, ChatGPT retains unstructured bits between conversations. A custom-built agent with persistent memory knows your projects in detail, the history of your decisions, and retrieves a voice note from three weeks ago on demand. It's not the same nature of memory.
Do you need to be a developer to live with an AI agent?
Not strictly. But you need to be comfortable with technical tools (terminal, APIs, deployment). I co-built mine with it: the agent often suggested what it needed to function better. I supervised, validated, corrected.
Is it risky to entrust your data to an AI agent?
The risk exists. My data goes through Anthropic's models. My rule: partial access only, never full access to my emails or my entire calendar. I'm working toward a more sovereign solution with open-source models.
Does an AI agent replace a note-taking tool?
No. It does something fundamentally different: it doesn't store, it connects. It links an idea from three weeks ago to today's goal. My notebooks never did that.