Think about what that actually means. Yes, the newest tools have started to keep a few notes on you. But it is a thin thing, a sticky note next to a real memory. Close the thread and most of it is gone. Tomorrow the model meets you half a stranger, running the same year-old job for the first time again. We keep making the model smarter, and we keep giving it a memory with a hole in the bottom. For the kind of work a person does over months and years, that is not a small annoyance. It is the ceiling on the whole thing.

A bigger AI is not a smarter one.

The obvious answer is, fine, just tell it everything. Paste in your whole history at the start of every conversation and let it sort things out. It turns out that does not work, and the reason is worth understanding because it is the crux of everything that follows.

An AI can only hold so much in mind at once. Researchers call that its context, and you would think the fix is simply to make the context bigger and pour more in. So they tested it. They hid an important fact in the middle of a long stretch of text and asked the model to find it. The model kept missing it. The finding got a name that says it all: lost in the middle. Pile on more and it does not get more reliable, it gets less. Newer tests keep landing in the same place. The size of the window is not the same as the amount the model can actually use.

A bigger context window is a bigger desk, not a better librarian.

Here is the way I finally understood it. Ask a good librarian a question and they do not wheel every book in the building over and stack it on your desk. They walk to one shelf and hand you the three books you actually need. The hard part was never having a bigger desk. The hard part is knowing which three books, out of everything ever written down, this exact moment calls for. That skill has a name in the field, and the entire industry is racing to build it right now. So when someone tells you AI memory is a solved problem, it is not. It is the frontier. It is the part still being figured out, this year, in public.

What it is like to be remembered.

Now flip it around, because the payoff is the whole reason to care. You forget your own working life too. Not the big milestones, the texture. Why you made the call you made back in March. The reason you dropped the second vendor. The lesson one project beat into you that would have saved you in the next one, if you had remembered it was there. That knowledge rarely gets written down anywhere durable, so it leaks away, and you end up solving the same problem twice.

An AI that actually holds your context, including the parts you have forgotten, changes the shape of your day. The next decision gets easier because it is standing on every decision before it. And this is not a hopeful guess, it is one of the better-documented findings in how minds work. When people know they can look something up later, they stop straining to hold it in their head and free up the room to actually think (researchers call it the Google effect). A reliable memory outside your head does not make you lazy. It hands you your working memory back.

I feel this one personally. I have an AuDHD brain, and the hardest part for me was never having ideas, it was holding the thread long enough to land one. An outside memory that keeps the thread while I chase the idea is the difference between a thought that becomes something and a thought that evaporates by lunch. What helps a brain like mine most turns out to help everyone a little. The rest of you are paying the same tax, just quietly enough to blame yourselves.

Now the uncomfortable part.

A thing that remembers your entire working life is powerful, and if that does not make you a little uneasy, you are not paying attention. Because the immediate next question is the one that decides everything. Who owns that memory?

Most workplace AI answers it the worst possible way. It remembers you for your boss. It watches how you work and reports upward, and people are not fooled for a second. In a national survey, 61 percent of workers opposed their employer using AI to track them, and two thirds figured whatever it collected would eventually be used against them. So people do the only rational thing. They feed the watcher nothing true. The tool starves on the exact signal it needed, and it dies quietly, and everyone blames the technology. It was never the technology. It answered the ownership question wrong.

The bargain that actually works.

There is a version where both sides win, and it comes down to one hard line drawn in the right place.

You own the memory. All of it. A plain, readable ledger of everything the system knows about you, where each piece came from, and a delete button that genuinely empties the drawer instead of hiding it. It is sealed and encrypted. And when you leave, you take it with you, the way you have always been allowed to walk out with your own notebook. That last part is not exotic. Europe already treats the right to carry your own data from one place to another as law.

Your employer never sees any of that. What the employer gets is the weather.

A forecast warns a whole city that a storm is coming this afternoon without naming a single raindrop.

That is exactly the right amount of information, and exactly the right limit on it. The institution can see that a department is straining, that burnout is climbing ahead of a wave of people quitting, that a team is heading for a capacity cliff. It sees all of it at the level of the group and none of it at the level of the person. And that wall has to be built into the plumbing, not written into a policy, because we have decades of proof that a policy does not hold. Researchers showed that 87 percent of Americans can be picked out of supposedly anonymous data using nothing but their ZIP code, birthday, and sex. Others re-identified real people inside an anonymous movie-ratings dataset from a handful of ratings. Stripping the names off after the fact is a promise that keeps breaking. So the honest design never gathers the individual into the aggregate in the first place.

The strange, good part is that the person's safety is what makes the forecast honest. People tell the truth to a system that cannot be turned against them. Watch people and you get performance, everyone acting fine. Protect them and you get signal, the real weather, because nobody had to hide from it. The guarantee for the person is the thing that makes the data worth anything to the institution.

What we are building at FirmSideAI.

Every tool we make started as one I needed and could not find. This one is called Cortex Engine. It is a working memory that belongs to the person, and a forecast that belongs to the institution paying to host it, with the wall between them built into the architecture instead of promised on a page. The individual version, just called Cortex, is running today.

Picture where it goes. An AI that grows with a person from school, moves with them through every job they ever take, and holds their context including the parts they have forgotten, so their skills compound over an entire career instead of leaking away every night when the day runs out. The model in your pocket will keep getting smarter on someone else's schedule, and that is fine. The memory is the part that should be yours. That is the part we are building, and that is the part we think you should get to keep.

For a long time I thought that was a personal failing. I have an AuDHD brain, so it was easy to file this under "one more thing I cannot stick with." I was wrong about that, and the numbers say so plainly.

You are not the one quitting. Almost everyone is.

Here is the part nobody selling you an app puts on the box. Across the whole app world, about three out of four people stop using a new app within the first three days of installing it. That is not my number. It shows up again and again across the big measurement firms that track billions of installs.

Productivity apps do a little better than average, and I want to be fair about that. The best of them keep somewhere between twelve and eighteen out of every hundred people around at the one-month mark, and the typical one keeps about eight. Read that the other way and it is brutal. Even in the category that is supposed to be good at this, roughly eight or nine out of every ten people are gone within thirty days. That figure comes from a 2026 benchmark study that looked at more than thirty-seven thousand apps.

So the drawer full of dead apps is not a you problem. It is the normal outcome. When something happens to nearly everyone, the honest move is to stop blaming the person and start looking at the thing they keep walking away from.

What actually kills the app is the upkeep.

I will be straight about the evidence here, because that matters to me. There is no single clean study that proves "manual data entry is what makes people quit." That research does not exist yet. What I have is fifteen years of standing behind a counter running my own shop, my own repeated experience with every one of these tools, and a stack of research that all points the same direction. So I am going to make an argument, tell you where it is solid, and tell you where it is me connecting dots.

Here is the argument. A productivity app is only useful if it is fed. It needs to know your tasks, your notes, your deadlines, your projects. And in almost every app on the market, the person who has to feed it is you. You are the data-entry clerk. You type in the task. You tag it. You move it to the right list. You update it when things change. You keep the whole thing current, by hand, forever.

That upkeep is the tax. And it is a tax you pay every single day, whether or not you got anything back that day. The app made a deal with you that most people quietly decline after a couple of weeks: do my filing for me, and maybe I will be useful. When the filing costs more attention than the app returns, you stop filing. Then the app goes stale, and a stale productivity app is worse than no app, because now it is lying to you. So you close it. Within a month. Like almost everyone.

My brain just hits that wall first.

This is where being neurodivergent stopped feeling like a weakness and started feeling like an early warning system.

The research on ADHD is clear on one thing. The hardest part is not doing the work, it is starting it, sorting it, and keeping it organized. That is what "executive function" means in plain terms: the part of the brain that handles getting yourself going and keeping yourself sorted. When that part runs differently, every bit of manual upkeep an app demands lands twice as heavy.

You can see it in the dropout numbers. When researchers build digital tools for people with ADHD and actually measure who sticks around, they find dropout rates running from about one in five to nearly half. One well-known training program only got 38 percent of people to finish, and when asked why, people said it took too much time and they lost interest. A 2026 review of 133 of these studies found that most of them either had huge drop-off or did not bother to measure it at all.

I am not telling you this so you feel sorry for the ADHD brain. I am telling you because that brain is the canary. It quits the high-upkeep tool first, and loudest. Everyone else is paying the same tax, just slowly enough to blame themselves instead of the design. My deficit is the differentiator here. It made a problem visible that everyone has.

And to be honest about the other side: when a tool is designed right for this brain, it works beautifully. In one research project, an app built hand in hand with ADHD users kept 97 percent of them coming back. Good design is most of the game.

Then there is the second tax: living in twelve tabs at once.

Say you push through and keep your app fed. There is still a second bill waiting, and this one hits everyone regardless of how their brain is wired.

Most of us do not use one tool. We use a pile of them, plus a wall of browser tabs, and we bounce between them all day. Every one of those little jumps has a cost that has been measured in a lab. In a classic study, after just 20 minutes of getting interrupted and switching around, people's stress and frustration rose by roughly a third, and the work simply felt harder to do. They actually finished faster, but only because they were pushing themselves harder to make up for the chaos, which is not a trick you can pull every day without burning out.

There is a second finding that names the exact feeling. It is called attention residue. When you jump off one unfinished thing to go do another, part of your mind stays stuck on the first one, and you do the next task worse for it. Gloria Mark, the researcher who has studied this longer than almost anyone, is widely cited for the finding that it can take around 23 minutes to get back to the task you were pulled away from.

So every app you add to the pile does two kinds of damage. It is one more thing to feed, and it is one more doorway you walk through all day, leaving a little of your attention behind each time. The pile itself is the problem. Which is why the answer is not a better app to add to the stack. It is fewer places to look.

What changes now: the machine can do the filing.

Here is the part that made me start building instead of just complaining.

The reason the upkeep tax existed is that software could not understand your messy human information on its own. It needed you to sort it into neat little boxes first. That is the thing that just changed. AI can now read the messy version. It can pull in the information a system needs on its own, look at it, decide what actually matters, and keep the whole thing current without turning you into its clerk.

We already have early proof this cuts the load in the right setting. When doctors got an AI that quietly listened and wrote up their visit notes for them, instead of typing everything in by hand, the mental workload of the task dropped by about 65 percent in the study that measured it. The filing did not go away. It just stopped being the human's job.

Now I have to be honest, because hype is the fastest way to lose you, and because I do not actually believe AI fixes anything on its own. That doctor study was small, it was early, and one of its authors worked for the company that made the tool, so hold it loosely. And when you zoom out to the whole economy, the picture is genuinely mixed. A 2025 review of 371 studies found no reliable across-the-board productivity boost from adopting AI. Another large review found that people working with AI often did worse than whichever one, the person or the AI, was better on its own, except on open-ended creative work.

I do not find that discouraging. I find it clarifying. AI is not a magic wand you wave over a broken tool. It is an equalizer or an amplifier depending entirely on who builds it and how. Point it at the part people actually hate, the endless manual upkeep, and design it around the real work, and it earns its keep. Bolt it onto the same old cluttered app as a chat box in the corner, and it does nothing but add noise.

What we are building at FirmSideAI.

Every tool we are making started as one I could not stick with myself. That is not a marketing line, it is the actual origin. I lived the fade-out over and over, and I got tired of blaming my own brain for a problem that was baked into the design.

So we build to two rules. First, the tool feeds itself. Wherever the machine can pull in and sort the information on its own, it does, so you are not the clerk anymore. Second, one place to look. We collapse the work into a single dashboard instead of scattering it across a dozen apps and a wall of tabs, so you stop paying the switching tax just to see where things stand.

None of this is about being impressed by AI. I am not. It is about a specific, boring, human problem. The tools meant to give you your time back have been quietly charging you for it, in daily upkeep and daily distraction, and most people pay until they cannot anymore and quit. That bill is the thing we are trying to cancel.

The app was never the problem. The upkeep was. And for the first time, the upkeep is something a machine can carry instead of you.

Which leads to the only question that matters: can you actually see that record before you fill in the box? Right now, mostly, you cannot. Not in one place. Not without wading through each campaign's own version of itself.

Why the Civic Analytics Lab exists.

First State Lens is the Civic Analytics Lab at FirmSideAI. The thinking behind it is not complicated. The public record already exists. Roll-call votes, campaign finance filings, official candidate lists, statements made on the record. It is public by law. It is just scattered across a dozen government sites, formatted for clerks instead of for people, and easy to lose under the noise of an election year. That is not a scarcity problem. It is a logistics problem, and I have spent my whole working life on logistics.

So we take the primary record, put it in one standardized place, link every single fact back to the exact document it came from, and then we get out of the way.

Why nonpartisan is the entire point.

Here is the rule we hold ourselves to. Every candidate for an office gets the same fields, in the same order. Campaign finance appears as the raw number the candidate disclosed, with no adjective attached to it. Positions appear as the candidate's own words, quoted and dated, never our paraphrase of what they supposedly meant. Where a candidate has nothing for a field, a first-time candidate with no voting history for instance, we say that plainly instead of leaving a blank that reads like something is being hidden.

We do not rank, endorse, or grade anyone. The moment we tell you what to think about a candidate, we have handed you our opinion wearing the costume of data, and you have lost the one thing that made the page worth opening. A voter looking at two names does not need me to decide for them. They need the record, in the same shape for each person, with a link to check every line themselves.

Why I believe in this.

I have come to think that most of what separates a good decision from a bad one is not raw intelligence or gut instinct. It is data and context. The verifiable facts, plus enough surrounding truth to read those facts correctly. I ran a small business for thirteen years, where a wrong read cost real money, and the pattern held every time. The people who did well were rarely the smartest in the room. They were the ones who could see clearly and were not talked out of the numbers by the story sitting on top of them.

Casting a vote is that same act. You are making a decision that carries real weight, and you deserve to make it with the record in front of you instead of a slogan.

What this tracks, and when it opens.

The first record we are publishing covers the race for Delaware's single U.S. House seat. For each candidate: their official filing, their votes if they currently hold office, their campaign finance straight from federal disclosures, and their positions in their own quoted words. From there it grows to the rest of the 2026 ballot, U.S. Senate, the statewide offices, the General Assembly, and county government.

The full record goes public the week of July 14. That date is not arbitrary. Delaware's candidate filing deadline is noon on July 14. Until that deadline passes, the field is not final. Some candidates have filed and others have not filed yet, and publishing a side-by-side of a ballot that is not actually set would be its own quiet form of distortion. A nonpartisan record has to show the complete field, or it is neither complete nor honest.

So we built it, we verified every line by hand against the primary sources, and we are holding it until the ballot is real. Then it opens. If you vote in Delaware, this is being built for you.

I want to be honest about why this one matters to me. I am not naturally good at long, complicated, manual processes. My mind does not hold forty pages of instructions in a neat stack, and for a long time I thought that was a personal failing. It took me years to see it differently. A lot of these processes are not hard because the work is hard. They are hard because nobody ever built them for a regular person with a regular amount of time.

Government contracting is the sharpest example I know.

The process keeps good people out.

Ask around and you hear the same three things. The bids feel wired, shaped in advance for whoever has the right relationships. The instructions are unclear in ways that only ever hurt the newcomer, because the insider already knows what the words are really asking for. And the whole thing runs on bottlenecks that neither side has the resources to fix. The small shop is drowning in unpaid nights. The contracting officer on the other end is buried under more notices than any person could write clearly.

None of that is the fault of the operator staring at the forty pages. It is a system that quietly filters for stamina and connections instead of for who can actually do the work. Good people get screened out before they ever get a fair look. That is the part that has always bothered me, and it is the part I think is finally starting to give.

The whole game runs on public data.

Here is what almost nobody selling you software says out loud. Most of what you would want to know before you make that late-night call is already public, and a lot of it is machine readable. SAM.gov lists the opportunities. USAspending and the federal procurement records show every award the government has made, to whom, and for how much. Past performance records show who delivered and who did not. Wage determinations set the labor rates. The solicitation keeping you up is itself a structured document, with the scoring written right into it.

None of it sits behind a paywall. The law says it has to be open. The federal government spent about 755 billion dollars on contracts in 2024, and it has to show its work.

The size of the big players is working against them.

The primes handle all that data the way big companies handle everything, with people. A capture team. Analysts who pull the records, read the budgets, and decide what is worth pursuing. That team is why it can feel like they see around corners you cannot.

But a team is a cost, not a magic trick. It has meetings. It has handoffs. It has a Tuesday where two of the analysts are out and the decisions do not get made. A capture team of twelve moves at the speed of twelve people trying to stay in sync. Their real weakness is that they need a capture team at all.

The judgment that team provides, whether a given opportunity is worth chasing, is exactly the kind of judgment that public data plus a little AI can now put in one person's hands. On the decision that comes before you write a single word, small no longer has to be slow.

The trick is writing fewer proposals, not more.

Most of the AI being sold to contractors right now promises to write your proposals for you, and to help you chase far more of them. I went looking for the evidence behind the "three times the opportunities, half the time" numbers, honestly hoping to cite them. They came apart in my hands. The strongest versions did not survive a careful check.

And chasing more was never the goal anyway. Writing is the expensive part. A real proposal is a week you do not get back, and most of that week, across this whole industry, gets spent on bids that were lost before they were written. The edge is not writing faster so you can lose faster. It is deciding better, so the only proposals you write are the ones you had a real shot at.

Decide first. Write second. And write a lot fewer of them.

What we are building, one day at a time.

That decision is what BidIQ+ is for. You give it a single opportunity and your capability statement, and it scores how well the two actually fit, from one to a hundred. If the score clears the bar you set, and only then, it will write you a first draft to build from. The judgment comes first. The writing comes second, and only for the bids that earned it. It turns the 2 a.m. guess into a number you can look at before you commit your week.

And the timing is kinder than it has been in a while. The government has started rewriting its own rules and pulling more than a thousand requirements out of the process, which means less to comply with on every bid. The simplified threshold, the lane where small firms have the natural edge, moved up to 350,000 dollars. Small businesses won a record 183 billion dollars in federal work last year, more than a quarter of everything spent, past the government's own goal for the fourth year running. The door is open wider than it has been in a long time. The only question left is whether you can tell, in time, which openings are really yours.

We are going to keep building for this corner of the world, and we are going to get a little better at it every day. Not because government contracting is glamorous. Because it is one of those places where the process itself has been quietly deciding who gets a shot, and I would rather help hand that decision back to the people doing the work.

The information you need has been public the whole time. The hard part was always making sense of it fast enough to act. That part, finally, is getting easier.

That is the gap. Right now, AI can go either way on it. It can hand even more advantage to the people who already have analysts, or it can put the same answers in everyone else's hands. FirmSideAI was built for the second one.

Data to the people, in language a person can actually read.

That is what First State Lens is. Every number shows where it came from. No page tells you what to conclude. We show you the data, you decide. When a number needs a caveat, the page says so in plain words, because the fastest way to lose your trust is to make a figure look cleaner than it is.

Built for the person who has to prove something.

If you spend part of your week proving something to a funder, a board, a reporter, or a room of neighbors, this is for you. The grant report that needs the real graduation number. The board deck that has to show where the money went. The op-ed that needs a figure that holds up. You should not have to be the well-resourced one to get a straight answer about your own government.

Open your government.

I am sitting in the Audion at the University of Delaware's STAR Campus. About two hundred people are in this room. Founders. Foundation directors. State representatives. University deans. A few of us building AI companies. More of us trying to figure out what AI is going to do to our work over the next ten years.

Sunita Chandrasekaran has been at the front, running the room with real precision. Krista Griffith, the Chair of the State AI Commission, spoke this afternoon. It is a good event. Delaware is doing real work on this.

And I keep hearing one question, asked five different ways.

How does this Institute make sure AI helps close the wealth and skill gap instead of widening it?

It comes from the front of the room and the back. From a foundation officer. From a small business owner. From a community college dean. From a young woman who introduced herself as a workforce development specialist. Every panel gets a version of it.

I want to write down what I think the answer is, because I think most of the conversation around AI equity gets the diagnosis wrong. I want to be on record about it.

The technology is not the gap.

Anyone can use ChatGPT. The model is essentially free. The tools are getting cheaper and easier every quarter. If technology access were the gap, the gap would already be closing.

What is actually the gap is everything around the technology. Training pathways. Customer access. Capital. Mentorship. Social proof. Time. Executive function. The connecting tissue that turns "I can open the app" into "I can build a business with this, get hired for an AI-supporting role, or restructure my work around it."

People who already had those things are using AI better. People who did not have them are using AI at the surface and falling further behind.

This means most AI literacy programs, taken on their own, will widen the gap they are trying to close. Literacy is the floor. It is not the ceiling.

The real leverage is downstream of the model.

If I had to bet on where the gap actually closes in Delaware, here is what I would watch.

State procurement. A meaningful share of state AI contracts going to small and minority-owned firms. A structure for AI work specifically, modeled on existing DBE programs. If state spending on AI goes 90% to enterprise vendors and 10% to working-class use cases, the gap widens. If that ratio inverts, it closes.

Workforce pipelines into AI-supporting roles. Annotation. Evaluation. Prompt engineering. Change management. Ethics review. Many of these are real wage opportunities that do not require a four-year degree. The labor market for these roles is already opening. The question is who gets routed into them.

Capital pathways for first-time founders. Programs like Delaware State University's Innovation Venture 2.0, which just funded Malcolm Coley's PrizeChips at $40,000. Pre-seed capital structured to find founders who do not already have the network.

Commission composition. AI Commission seats. Advisory boards. Procurement rule-making bodies. If the population most affected by AI's economic shift is not at those tables, the rules will optimize for someone else.

Civic AI that does not pre-suppose digital literacy. Voice-first interfaces. Accessibility defaults. The DMV portal pattern Delaware is already building. Public services where AI is in the background making things easier, not in the foreground gating access.

Five things. None of them are about teaching ChatGPT.

The trap is the constituency that shows up.

This is the part that gets missed in every AI equity discussion I have been in, and it has nothing to do with intent.

State AI institutes default-serve the people who show up to events. Foundations default-fund the projects they hear about. The constituency most at risk of being left behind by AI, by definition, does not show up to AI summits. They are working, caring for kids, trying to hold onto their current job. Flying to Newark on a Tuesday to attend a State of AI panel is not on the calendar.

If the Institute waits to be asked, it will widen the gap by default. Not because anyone in this room wants that. Because the structure of who shows up determines whose problems get worked on.

Equity has to be designed in. It does not emerge.

Why this matters to me specifically.

I built FirmSideAI on the thesis that AI is at a fork in the road. It can be the largest equalizer of this century. Without intentional work, it will instead amplify the inequality that already exists. Which version it becomes is not a question of technology. It is a question of who decides to make it the equalizing version, again and again, in rooms like this one.

I am also someone who was late-diagnosed neurodivergent. I spent fourteen years running a small retail business with an operating system in my head that the world did not understand. AI, properly built, is the digital prefrontal cortex that the neurodivergent population needs. Not as a metaphor. As a tool that does the executive function the brain underweighted, so the parts the brain overweighted finally land in the world.

That gives me a specific lens on the equity question. The population that benefits most from properly-built AI is the population that has been hardest to reach with conventional support. They are not just underserved. They are structurally invisible to most of the systems that decide where money and attention go.

If a state AI institute is serious about closing the gap, that is the population the work has to start with. Not as a side program. As the center of the program.

What I am asking the Institute to commit to.

Pick one underserved population in Delaware. Design an entire AI uplift program around them, end to end. Capital access, training pathways, customer connections, advisory representation, the full stack. Make it a real five-year pilot with measurable outcomes. Report publicly every year on whether it worked.

The Institute earns the right to lead on equity by doing one thing well. Not by mentioning the word in every panel description.

I will help. I am sitting in the room. I have a company built around this thesis and a network forming around it. If anyone reading this is in a position to scope that pilot with the Institute, talk to me. We can start in this room.

The point is not the post.

I am writing this from the back of the Audion because I think a question worth asking five times is a question worth answering once. The most useful thing I can do today is name what I think the answer is, in writing, while the people who can act on it are still in the same building.

If you are at the Summit and want to talk through any of this, find me before the closing remarks. I am wearing the FirmSideAI badge.

Mark Sanders is the founder of FirmSideAI, based in Newark, Delaware. He builds AI infrastructure for neurodivergent operators and civic-tech applications. He is reachable through firmsideai.com.