AI That Remembers Everything Told Us "Overloaded" Every Morning

An AI assistant that remembers everything is the easy part. Point a model at your email, calendar, call notes and task list, and it will remember more than you ever could. The hard part is deciding what it tells you at 6am. Ours remembered every open task my co-founder and I carried, scored all of them against a six-hour day, and told us “12.75h of real work, overloaded” every single morning. The memory was fine. The morning message was useless, and it took us eight days to see why.

The digest

In September I built a small job that posts to our Slack at 6:00 Eastern, Monday to Friday. It reads our open tasks and runs them through a Tim Ferriss-style 80/20 filter: eliminate first, then automate, then delegate, and only then do. The idea was simple. Before the day starts, tell each of us what to cut or hand off, and what’s left that only we can move.

I put real care into the writing. The prompt asks for a sixth-grade reading level, one idea per sentence, and bans a list of jargon, because the message has to be readable before coffee. There’s a size test that fails if the post grows past a few short blocks. It went live on September 7.

The first live check already caught one bad number. The workload measure was counting the whole company’s board (432 active tasks) and calling it the founders’ load. Scoped to the two of us it was 246, and our “firefighting” share (work created and closed inside 48 hours, so it never appeared in a plan) dropped from 29% to 8%. Had I trusted the first version, it would have accused us of firefighting we weren’t doing.

The bigger problem was sitting in plain sight.

A number that never moves

The digest scored every open task, about 230 of them, against a six-hour workday. The answer was the same every morning: 12.75 hours of real work, overloaded.

That was true, in the sense that 230 tasks is more than a day. It was also the same sentence every weekday. A warning that never changes stops being information. You read it once, you nod, and from then on your eyes slide past it to whatever’s underneath.

This is the trap with any AI that remembers everything. The more it remembers, the more tempting it is to report all of it. “You have 230 open things” is accurate and impossible to act on. I’d already seen the same pattern from the other direction, when the brain created 1,042 tasks in four months from our email and meetings. Remembering more had made every list longer, and longer lists were getting read less.

What changed on September 15

The fix was to stop analysing the backlog and analyse the day. The digest now reads today’s slate: anything due today or overdue (an overdue task still has to happen today), plus Urgent tasks with no date at all, since in our system Urgent means today or the next business day and a missing field shouldn’t hide the loudest thing on the board. Everything else is still counted, as one line of context, and never as the headline.

Running it against the live board that day:

• 117 of 231 open tasks were due or overdue. That’s a real number about a real day, and it’s alarming in a way 12.75 never was, because it changes as we work.
• 44% of that slate was missing a project, a type or a due date. When the board is that messy, the digest says its read is a rough guess instead of printing confident analysis over bad data.
• Firefighting sat at 10%, under its threshold, so that alarm stayed quiet.

One small change mattered more than I expected. The old digest, given an empty list, said “no open tasks, or the run failed.” The new one says “nothing due today.” If your assistant treats a good morning as a possible error, you learn to read every message with suspicion.

The button that remembered nothing

We’d learned a version of this a month earlier. Our Friday time-log bot posts a table of who has logged hours that week, then checks again at 16:30 and tags anyone with nothing logged. From August 7 to September 4, each row had a Confirm button.

Clicking it rewrote the row to say who confirmed and when. It stored nothing anywhere. Nothing downstream read it. The 16:30 check recomputed from scratch and still tagged people who had confirmed a zero. Worse, the table went out at 9:15 on Friday morning, before anyone had logged Friday, so we were asking people to confirm a table guaranteed to be incomplete. It was a ritual with a button on it. We removed it.

What replaced it is a rule I now apply to anything that reports on people: three buckets, never two. Logged, zero, and not verified. If the bot can’t match someone cleanly (a duplicate row in the team database, say), it names them without tagging them and flags it as a data problem. A check that fails stays silent. A check that’s wrong tags someone in front of their team for hours they actually logged, and that costs more than the silence.

That same instinct killed a check before it shipped. We specced “you closed a project this week and logged nothing against it,” then looked at the data. Only 15 of 33 completed projects had an end date, and where they did, it disagreed with the last time entry by as much as 61 days. It would have fired zero times in two months and then fired wrong.

If you have ADHD and you’re picking a tool

Some people look for an AI that remembers everything because keeping track of commitments is genuinely hard for them, ADHD or not. If that’s you, everything above applies with more force, because a tool that nags you with a number you can’t change is one more thing to feel bad about and then ignore.

Here’s what I’d want from an assistant that remembers for you, based on what broke for us:

• It should show today, and keep the backlog as a count. You want “three things are due today, here they are.” You don’t need a daily verdict on your whole life.
• Every item should come with its source. “You said you’d send the quote Friday” is only useful with the email or call line attached, so you can check it in one click instead of wondering if the AI made it up.
• It should say when it isn’t sure. “I couldn’t tell who owns this” is better than a confident wrong answer. Ask any tool you’re evaluating what it does with messy data.
• A clear day should look clear. If the tool’s empty state reads like an error, you’ll stop trusting the full state too.
• It shouldn’t ask you to maintain it. Confirm buttons and daily check-ins are the first thing to fall apart in a busy week. Our Confirm button was live for four weeks and recorded nothing.

What it means for the rest of us

Memory is cheap now. Our company brain reads our email, calls and docs, and in four months it surfaced 626 threads where we owed someone a reply. None of that helps if the morning message says the same thing every day.

The test I use now is whether the message could have been written yesterday. If it could, it’s wallpaper. That’s the same lesson as the feedback loops that collected nothing: a system that runs every day and changes nobody’s behavior is decoration, however well it’s built.

If you want this for yourself, the ADHD page shows what it looks like in practice: you ask “what did I promise this week?” and get the answer with the email or call it came from, built on the email, docs and calendar you already have, so there’s nothing to log. The cheapest way in is a 45-minute coaching session for $99 where you do the typing over screen share, on your own documents, at your own pace.