In about eight weeks, Eric Edouard built an AI that remembers his whole life (every conversation, message, email and photo) and agents that run a shop's website, ads and content on their own. Here is what it does, in plain words.
He only wore a small recorder. Nobody took notes. Ten minutes before, he asked out loud what to check and got three precise points from earlier conversations.
122,500 emails, about 270,000 chat messages, 41,900 photos and every spoken conversation since September, all in one memory.
Built by an agent working toward a goal, then improved through about 160 reviewed changes in 15 days.
The agent noticed that 47 forgotten bids were eating 76% of the budget, asked for a yes, fixed them, and checked the result.
It reads the till, website analytics, Google search and ads, then emails a ranked list of what to do today.
1,720 changes, written with AI and reviewed by AI. still a work in progress, used every day
Eric Edouard, software engineer, French and Thai, based in Bangkok. He co-founded MADworks, an AI automation agency, with his partner Rémi.
A two-hour meeting, five people, three languages. Instead of taking notes, he clipped on a small recorder and afterwards asked the AI: what was decided, what is open, what is blocking? It just worked. So he asked: what else does my life already say?
New messages are copied within seconds. Even messages someone deletes "for everyone" stay readable, and every old version of a note is kept.
Every night, AI reads the day the way a sharp assistant would, and checks every fact against the exact words it came from.
Every source is copied as it happens. Nothing is typed in by hand.
Each day is cut into conversations, threads and moments, across all apps.
What happened, where, with whom. Promises, plans, to-dos, each with its exact quote.
One person across WhatsApp, mail and voice. Their story with him, month by month.
Ask anything. It fills the calendar, the to-do list and wakes the agents.
The memory plugs into ChatGPT and Claude, so the assistants he already uses know his life, with a source for every answer.
Every night it works out what is planned, what is promised and what is late, and puts it where he already looks.
Said in a call, sent on WhatsApp, confirmed by mail: it becomes an event. A later ticket moves it instead of duplicating it.
What's new, what's due, what he promised and to whom.
Things he said he would do, in any language, kept until done.
When unsure, it asks one short question in chat, never a form.
A nickname on Instagram, a number on WhatsApp, an email: one person.
All taken from his own memory, with dates.
A display window at the shop broke. He sent the measurements, chased the quote by voice ("remind me to follow through on the glazier"), and when the pane arrived 4 cm too wide, the memory had the order, the invoice and the delivered size. The remake request was precise in one go.
The form photographed by his shop partner, the carrier's follow-ups, each step tracked as "waiting on someone" with the shipment number.
A lawyer's Thai messages, a recorded Thai conversation and a photographed Thai invoice became English to-dos with deadlines, the original quoted.
From his calendar, chats and recordings he found that 30 to 50% of his working time went to a family building project. Hard numbers to set priorities.
The cancellation, the rebooking and the claim, kept as one trail without filing anything by hand.
Earlier calls and notes fed an audit of a small salon's online presence: sent 31 Oct, approved 1 Nov, new website handed over 4 Nov.
A ride is attached to the trip, a fee to the form it paid for. No receipts typed, and a daily budget it can hold him to.
A loop is an AI agent given one goal it can never fully reach, like "become one of Europe's main online adult shops". It doesn't finish and stop. It plans, works, measures, learns, and goes again, day and night.
Lynx is his family's adult boutique. In six weeks, agents relaunched its advertising, rebuilt its search presence, built a full online store and started its social media again. Eric approved; the agents did the work.
In its weekly review the agent saw 47 old bids overriding the plan and taking 76% of the spend. It asked, fixed them the same day, and measured a week later.
Average cost per click on search. It then moved budget from the costlier campaign to the cheaper one, step by step.
The shop ranks #1 on its main local search: 22,734 impressions and 2,854 clicks in the audit window.
It also prepares the shop for AI search (ChatGPT, Perplexity), and checks every page for words Google's ad rules ban.
Reads the till, Analytics, Search Console and Ads every morning.
When a change broke the filters on the shop's listings, the agent caught it, fixed it within hours and added a rule to its own instructions so it can't happen again.
No listing change ships without clicking a filter live.It takes a competitor post that worked, remakes it in the shop's voice, and lets an AI judge pick the best versions. He only votes A or B. When he disagrees with the judge, the judge's rules are rewritten until it thinks like him.
Illustration with example titles. Rounds and agreement figures are real.
Now a standalone tool that can run the same way for any brand.
A memory of everything is only useful if it is safe enough to trust with everything.
At MADworks we build the same for teams: a memory across the tools you already use, agents with a goal and a number, and a human yes before anything leaves.