The Next Breakthrough in AI Coding Won't Come From a Better Model

AI coding tools are becoming ridiculously capable. A developer can describe a feature and get working code in minutes. A founder can build a prototype over a weekend. Someone with very little programming experience can create something today that would have needed a small development team a few years ago.

Yet there is one problem I keep noticing: our AI tools are getting smarter, but our projects still have poor memory. I think this will become one of the biggest problems in AI-assisted creation.

Every new AI conversation starts with a tax

Imagine you have been building a product for three months. During that time, hundreds of decisions have been made. Why did you choose PostgreSQL? Why was one API rejected? Why is authentication structured in a particular way? Which features were postponed? Which bugs have already been investigated? What did the customer ask for last week? What should be built next?

A human team gradually absorbs this information. AI usually doesn't. Open a new conversation and you start explaining the project again. Move from one coding tool to another and you explain even more. Change models and part of the accumulated understanding disappears.

We have somehow accepted this as normal. I don't think it will remain normal for long.

The future will involve many AI tools

People are unlikely to use one AI system forever. Today you might prefer Claude for one task, Cursor for coding tomorrow, ChatGPT for research or planning, another model that suddenly becomes excellent at debugging, a specialised agent that handles testing. Something that doesn't even exist today could become the preferred development environment twelve months from now.

That is healthy — competition is moving AI forward incredibly quickly. The problem begins when the intelligence surrounding your project belongs to the tool instead of the project. Your code can move. Your files can move. Your repository can move. Your project knowledge should be able to move too.

Think about what happens when AI actually remembers the project

Imagine opening a different AI coding tool tomorrow. You don't spend the first twenty minutes explaining your architecture. The AI already understands the important decisions. It knows what has been completed, what was rejected and why, your conventions, the current priorities, and which instructions apply to this particular project. It can see the relevant history without loading six months of irrelevant conversations. You simply continue working.

That sounds like a small improvement. It isn't — it changes how we work with AI.

Context becomes part of the project

Most projects today consist of code, files and documentation. AI changes that definition. A modern AI-assisted project also accumulates context: decisions, reasoning, instructions, patterns, lessons from mistakes, architecture choices, customer requirements, approved ways of doing certain tasks, things the AI should never do, and things it should always check.

This information is becoming almost as important as the code itself. Yet much of it currently lives inside chat histories scattered across different platforms. That feels temporary. I believe future projects will carry their own structured context with them — the AI tool will connect to the project, understand what it needs, perform the work, and leave behind useful knowledge for whichever AI comes next.

This could dramatically reduce wasted AI work

Anyone using AI heavily for development will recognise this experience. The AI solves a problem. A few days later, another conversation investigates the same problem from scratch. Or an agent changes something without understanding a decision made two weeks earlier. Or you spend several thousand tokens explaining architecture that another model already understood perfectly yesterday.

Better project continuity reduces all of this. The benefit isn't simply convenience — it means less repeated explanation, fewer contradictory decisions, faster onboarding of new AI agents, easier switching between models, better documentation as the project evolves, clearer visibility into what humans and AI changed, and more reliable execution of project-specific rules. The cost of AI may continue falling. The cost of lost context may become much more noticeable.

Coding is only the beginning

The same idea applies far beyond software. Consider a marketing team developing a brand: an AI has already learned the company's tone, rejected phrases, target audience, positioning and previous campaigns. Why should that knowledge disappear because the team decides to use a different model?

Consider someone writing a book. The AI understands the characters, timeline, writing style and earlier editorial decisions. That understanding should belong to the manuscript, not whichever AI happens to be open. The same applies to product development, research, consulting, design, legal workflows and business planning. AI-assisted creation is producing a new kind of asset — machine-usable project knowledge. We haven't fully worked out how to manage that asset yet. We will.

Skills may become portable too

Today, we repeatedly teach AI how we want work done: how to review a document, how to analyse competitors, how to create a release, how to test something, how to make a particular kind of decision. Over time these instructions become refined. They effectively become skills, and those skills should be reusable.

If I have spent months developing an excellent workflow for reviewing product strategy, I should be able to use it with whichever capable model I choose. The same applies to coding standards, research processes, testing procedures and creative workflows. This would create an interesting separation: AI companies provide intelligence, projects provide context, and users provide their preferred methods and skills. Each can improve independently.

Trust will matter as AI agents become more powerful

There is also a less exciting subject that may become extremely important: control. When AI mainly answered questions, mistakes were annoying. Agents can now modify files, write code, call tools and increasingly take actions. That makes traceability important — what changed, which AI changed it, why, whether the information was suggested or approved, and what the previous state was.

As autonomous systems become more capable, serious users will want a clearer record of the relationship between human decisions and machine actions. The AI workspace of the future will probably need some form of governed memory rather than an endless pile of chat transcripts.

The winners may change. Your project shouldn't.

Nobody knows which AI model will dominate three years from now. That is exactly why I wouldn't want years of accumulated project intelligence trapped inside one platform. Models will change. Coding environments will change. Agents will change. Prices will change. Capabilities will change. Your work should survive all of them.

I think we are moving toward an era where the most valuable AI infrastructure may be surprisingly quiet. It will sit between our projects and whichever AI we decide to use. It will preserve context, carry knowledge forward, apply the right skills, keep a trustworthy history, and allow us to change AI tools without effectively starting the relationship again.

We have spent the last few years asking "Which AI is smartest?" The next question may be much more important: "Who owns everything the AI has learned about my work?" The answer should probably be: you do.