We quickly realised that keeping some state across sessions has value, and many memory implementations have sprouted. But there is another way, and it unlocks a lot more than context.

Task first, session second. Or is it really a session?

One direction of integrating AI remains underexplored: reversing the usual order of creating the session first and then specialising it for a task. The task comes first, and there is exactly one persistent AI session attached to it. Once that session is made persistent, it can become a lot more than an AI session. Task Matter calls this larger scoped workspace a “Matter.” Each Matter can contain multiple persistent AI threads, with each thread dedicated to one purpose within that Matter.

Now, the above might make you wince a little bit. Does this mean we need a plan for everything now? Even if they are the right thing to do, plans are a pain to create and maintain. There is no question about it. The answer to making plans is a combination of

• AI,
• and being smart about it.

AI helps a lot in creating and maintaining the plan. “My wife, my two boys, aged 9 and 11, and I are going to France for three weeks. Help me plan it.” “The holiday needs to be postponed by a week. Update the plan.” These kinds of things. Not all items are relevant, so clean-up is necessary, but it will add steps where I think “yep, didn’t think about that”.

And “being smart about it” is allowing flexibility to the idea of 1 task = 1 Matter. The model which I found works best in practice is:

• 1 global Matter, which is also the Matter for the “Focus” area. That’s where you run regular syncs, create goals, and ask ad hoc questions.
• 1 Matter per goal: the highest level in a plan.
• 1 Matter per task: the level below a goal.
• 1 Matter per action item: the level below a task.

In practice, you always have a space where to work in. I tend to use the task level most. Goal level is great for re-organising the plan. Action item level can be handy when there’s a detail I want to query to death.

So what do we get with this persistent Matter per scope? Because the Matter isn’t disposable, you can put things in it! Besides AI threads, we can store task descriptions, attachments, and notes in the Matter. That’s what Task Matter currently supports.

From workspace to platform

The above design opens up a very powerful capability. Remember that AI knows the purpose of the Matter before even starting. This means that for every kind of work, we can have a corresponding plugin that’s automatically loaded. It is like software that recognises a file with a known extension and adapts the UI and behaviour accordingly. The available plugins are installed by the user from a marketplace. In Task Matter, I want to call these plugins “Matter Apps.”

A Matter Apps marketplace is what transforms Task Matter from a task-management app into an extensible Goal OS, one that can theoretically adapt to any set of goals. The Matter is an environment in which any sort of application can run and any sort of integration can happen.

Let’s take Coding Matter, the Matter App for coding currently in development. It is inspired by similar products like Herdr. Integrated with Task Matter, it maintains a 1:1 mapping between the plan action items and worktree branches. Implementation status gets constantly updated in the Task Matter UI. If findings warrant plan changes, these are brought to Task Matter as well. The user can connect to a coding session anytime to direct the implementation more closely if needed.

As another proof of concept, I want to target a kind of goal I tend to struggle with. Completing a DIY project is such a goal. A DIY Matter App could support modelling a plan based on a picture and measurements, giving a list of tools, material estimates, and covering safety guidance. Though to be fair, even without the dedicated Matter App, Task Matter does well on DIY projects.

These are two concrete examples, but the category and variety of apps that can be created has a similar breadth as what the App Store has to offer. Think learning apps, for subject matters, languages, music. Think admin apps, for taxes, company creation, government procedures. Think traveling apps, for itinerary planning, local history, kids entertainment. Behind all these apps are goals which are series of steps, with each step benefiting from external feedback, help and prioritisation.

Why this matters

Task Matter’s aim is to become such an AI task-management system, and I’m hoping other projects also go in the same direction. Today’s AI chat is fundamentally aimless, it exists to answer whatever question happens to come next. The task-first paradigm gives AI a purpose, persistence, and a workspace. Now we have concrete matter on which we can build something.

So there it is, Task Matter Limited is now an Irish company. No more stealth mode.

For those just joining: Task Matter is an AI task-management app I’ve been building. Its focus is helping you reach your goals by tying persistent AI sessions to a plan rather than leaving them as throwaway chats. I’ve written about the preview, the origin story, and the why in earlier posts.

Task Matter also has a LinkedIn company page now so, really, it cannot be more official than that.

Another thing I have been doing is sweating over a landing page where you can register your interest and get a chance at early access. There’s a little note field at the end that asks for a goal you might use the app for, but really, put anything you like: feedback, a quick hello, a rant about task management. Check it out here.

If you’ve been following along, thanks for the support. And if you’re new here, a warm welcome.

My aha moment for ChatGPT was when I was in a not-so-different situation where I had to cook a complicated dinner by a certain deadline. I asked ChatGPT to give me a step-by-step plan to have the dinner ready by a certain time. And I asked it to use “The Wolf” tone. Chopping steps, cooking steps, clean-up, everything was minuted, in the right order. Some nuggets:

Assumption: It takes about 1 hour 15 minutes total to prep and cook this dish.
That means we’re starting at 4:45 PM at the absolute latest. But I like buffer. We’re starting at 4:15 PM sharp.

or

Step 1: Prep (4:15 – 4:30 PM)
Chop, dice, measure, line up your ducks.

• Peel and chop onions, garlic, carrots, celery — whatever the recipe calls for.
• Rinse lentils.
• Check sausages. If they need pricking or splitting, handle it now.
• Measure out stock, open cans (tomatoes, etc.), prep herbs.
• Have a clean workstation, trash bowl, and tools at hand.
• …

also

… Backup Plan (if you’re behind)
If it’s 4:30 and you’re just starting, cut veg smaller to cook faster, crank up heat on the simmer (careful), and keep the sausages sliced for quicker browning.

You’ve got this. Keep cool, stay sharp, and don’t let chaos into the kitchen. If anything goes sideways — remember: I’m Winston Wolf. I handle it.

And there was even a bit of extra time to enjoy a cup of coffee. The Wolf does enjoy a good cup of coffee.

We all have goals we think are out of reach. I have a bunch. Building a business is the obvious one. But even seemingly trivial goals like using all these random veg from the veg box into a recipe the kids won’t throw at my face. Or repairing that rotten door frame crumbling away. Or the one that for some reason is invisible to some men like me, lowering the mental load needed to organise the kids’ activities and holidays. My wife is saying to her colleagues that I’m building an app to lower the mental load of organising everyone and everything. And everyone finds this amazing. Besides all the things she already does, she’s also a brilliant marketer. So we have all these goals, and what we really need is a Winston Wolfe in our lives. What I’ve realised is that AI can be the next best thing.

Claude Code, more than Codex I find, is particularly good at this “recovering”. Even if, really, it is a major blunder that needs a much deeper fix. It’s like a cat that will break all your lovely cupboard deco, ruin your curtains, and fall back on its feet and cuddle with you after being shouted at. And then of course you have to feed it.

The main constraint cats and AI have in common is they have to be quick. They rely on signals they recognise to quickly make a decision, and act. Understanding the full context is never an option. Falling back on your feet and continuing to be fed becomes the critical skill.

Anyways, it’s obviously not a fair comparison. I eventually got my cat to re-arrange all my deco and replace my curtains. It even re-built a new house, with extension, and 4 bathrooms and no kitchen. While accidentally setting my car on fire.

Now, there are many reasons I like “Task Matter”. Most of these reasons revolve around the word “matter”. It’s a wonderful word with so many meanings, and all of them are positive and quite meaty, while staying general. A very powerful word. “matter” can be synonymous with “substance”, as in “dark matter”. This is the most intended meaning here, where “task matter” implies that we want to completely describe and understand the task substance. Another meaning is the one used in the title of this blog “I matter”, as in “I count for something”. The counting-for-something meaning is perfect for an app that helps you decide if a task should be worked on today. There are variations around that second meaning. For example think of “be of significance”, “care about”, “relevant”. And finally, there is the meaning of “matter” as “topic”, as in “expert on the matter”. Again, a meaning that fits well as Task Matter AI is an expert on the topics a task covers.

See, no sign of pun around Taskmaster in there. But if that helps you remember the name Task Matter, I’ll take it.

So yes, this is partly a stack overview, and partly me retroactively justifying my choices now that they have become real product architecture.

The short version:

• Language: Scala on the server, Scala.js in the browser
• Build system: Mill
• Frameworks: ZIO on the server, Laminar in the browser
• Persistence: Postgres for profile and auth data, S3 objects for user data
• Browser persistence: IndexedDB
• Data sync: custom git-like commits
• Data format: RDF N-Quads
• AI orchestration: custom AI harness
• Deployment: one simple VPS for everything, for now

The stack is unusual, but it is not random. There are reasons and benefits behind each choice.

Scala, server and browser

Scala is probably the most controversial choice. For web projects, it is now a niche language. I know this is not the default startup stack. I also know future me may one day curse present me for making hiring harder. But for now, I still think Scala holds its own.

The first reason is Scala.js. It works very well, and it lets me share code between the server and browser. That matters a lot for an app where the client is not just a thin UI, but a real local-first application.

The second reason is the ecosystem. Scala’s ecosystem is smaller than it used to be, but the quality is high. Because the language has advanced features, it tends to attract engineers who enjoy thinking carefully about abstractions.

The third reason is stability. Once on Scala 3, the ecosystem feels relatively settled. Because the community is smaller now, established libraries tend to remain established for a long time. ZIO and Laminar have both been relevant in the community for years.

And finally, there is the type system. This is the classic Scala argument, and it still applies. Types reduce the surface area for mistakes and document intent. Good for humans and AI. There is also some TypeScript in the codebase, and I haven’t found that AI makes fewer mistakes or produces better abstractions with TypeScript. I haven’t gone into the rabbit hole of measuring that, though.

Mill, ZIO, and Laminar

I started with sbt and moved to Mill. The practical reasons were worktrees and monorepos. sbt-git had issues with my worktree workflow, and Mill has better support for partial builds and aggressive caching. This is not glamorous, but build tools become very emotional once they get in your way.

ZIO was new to me for this project. I picked it for the batteries-included approach: there is usually a ZIO sub-project for most infrastructure needs, and the same approach extends nicely to custom code. Server-side code isn’t where most of Task Matter’s logic is, though.

The browser side is where it matters. I chose Laminar partly because it is used by the Scala.js creator, and partly because it promised to handle complex shared state. In Task Matter, many UI elements reflect the same underlying state in many different parts of the app. Laminar makes those updates feel instantaneous, and that has been one of the pleasant surprises of the project.

Offline-first data

Task Matter is offline-first for both reads and writes. The client keeps recent history locally. The server keeps the authoritative log.

Updates are stored as git-like commits. Whether a user or AI changes a plan, the change history is available for review, undo/redo, reporting, or AI context. That matters because plans in Task Matter are meant to adapt constantly to reality. A plan that cannot absorb reality is not much of a plan.

RDF N-Quads are the most exotic part. The benefit is generality. A commit is just a set of added and deleted quads, plus metadata. That makes applying sync updates independent of the application data model. And it’s conceptually simple: the data is a set of quads, changes are lists of REMOVE and ADD quads, commit metadata is expressed as quads, and the schema is made of quads. Everything is a quad.

I’m not really using RDF to its full power, though. When I chose RDF, I also had in mind AI writing its own RDF to store and index context and knowledge related to plan entities. I haven’t abandoned the idea. There is just so much to do!

On the storage front, for each user, Task Matter maintains a current-state file and a log file. The server sends commit deltas, keeping sync quick. When the server has moved on, the client rebases pending local commits onto the server state. Valid commits get new parents; redundant or invalid commits are dropped and surfaced to the user.

RDF sounds a bit strange, but in practice it gives me exactly what I wanted: offline edits, understandable history, and a data model that can evolve without rewriting the sync layer.

AI orchestration

AI orchestration has been the hardest part. Compared to the git-like sync, it is another level of difficulty. Task Matter always runs AI in a known context: a focus area, a goal, a task, or an action item. That creates a real opportunity to minimise turns, tool calls, and context size by tailoring the interaction to where the user is in the plan.

Finding the right abstraction took a while. My first attempt was multi-agent orchestration with specialised agents. It sounded right, but it was non-deterministic and impossible to debug. It was the wrong abstraction. Claude Code and Codex were not especially helpful there either. They can implement any design you give them, and they will even compliment you on it (especially Claude), but they’re as clueless as I am about agent orchestration.

What worked better was a set of built-in directives, each with its own recipe for system instructions, context, and available tools. A quick, cheap resolution step chooses the directive from the user prompt. The result is more boring and less magical than the multi-agent version, but it gives me much more control.

In that sense, Task Matter is a bit like a coding AI harness, but for goal plans.

To conclude

I have used Claude Code, Codex, and more recently Pi Coding Agent to write the code. I tend to give small tasks and adjust the results a lot after the fact. Even if the tech stack is unusual, AI hasn’t struggled one bit with understanding what the code does or writing code that compiles on the first try. At one stage, I was trying to find excuses to migrate to a more mainstream stack. I even tried to one-shot the rewrite of the codebase into TypeScript and React. The result was terrible. It made a completely different app that looked like the thousands of project management apps out there. Most of it didn’t work anyway. At least with an unusual tech stack, AI can maybe forget some of its learned UI patterns and let itself be pulled towards something more original.

What I wanted is an app whose sole focus is to help me get where I want to be, while having access to that AI knowledge. I also need to be able to manually interact with the plan. It’s my plan after all. One way to describe what I want is a ChatGPT focused on my goals and tasks. ChatGPT or Claude do integrate with project management apps, but really, the friction of integration is significant.

AI chats have been incredible to help me tackle difficult goals. They just need that bit of structure to help with keeping the focus to actually reach these goals. In Task Matter, I settled on a structure of Goals that contain Tasks that contain Action Items. To organise my week and days, I, or AI, bring the most relevant tasks into a focus area. But AI is there at every scope: global focus, each goal, each task, each action item. One AI feed for each scope. And that feed is accessible whenever you click on the object. The feed is a lot more than an AI chat, but that deserves its own post. I am very very pleased with the feed. Lots of eureka moments to get there and make it just work.

That went a bit further into the “how” than I first meant. I can’t wait to write more about it. Coming back to what this is all for, Task Matter’s ultimate objective is to help users tackle any goal, using best available knowledge for each one. As a realistic first milestone, Task Matter is focusing on mid- to long-term personal goals. For me, it has been building Task Matter itself, company creation, GTM, personal admin, DIY, running, traveling and gardening. It is all over the place, but it is simply what is filling my life at the moment.

While creating a draft for the landing page, I’ve realised I needed to spell out the answer to “why” and “what for”, which I’ve answered above, with some “how” leaking in. I have no idea yet how I will summarise the above in a couple of sentences. Though I have some cheesy product taglines ideas already. Let me know what you think:

• My favourite one, “AI with purpose, your purpose”
• One stolen from the tennis coach Patrick Mouratoglou, “Don’t think you can’t, wonder how you will”
• Variation on that theme, “Assume you will, discover how.”

If you have ideas to suggest in the comments, be warned I won’t pay you if I choose them! Though I might give you attribution :)

My story is both banal and existential. Like many in the big tech sector, I burned out. I didn’t understand at first what was happening to me. I was not over-worked, enjoyed the technical challenges, and had great colleagues. I understood later that my burnout was a little more complex. It was due to a disconnect between work I thought mattered, and reality.

My employer at the time was LinkedIn, though the employer itself is irrelevant, the context did matter. I was a software engineer in a big tech company, working in a satellite engineering office here in Dublin, while most of engineering was located in the SF Bay Area. Following a reorg and team change, my colleagues, projects and org dynamic had all changed. I found myself working on a core service used widely across all systems. The online service ran reasonably well. My focus was on an area that desperately needed investment, the data ingestion flow. The thing about data ingestion is that it’s never a blocker. Even if the data is stale, systems continue to run. Restarting sustained investment and getting buy-in on a data ingestion project is not easy. It needs proper lobbying, ROI discussions, and technically challenging design work. The technically challenging work did appeal to me, but not the lobbying and ROI discussions. When meeting SF-based colleagues from Dublin, these discussions compete with personal time. Up to that point in my career, I had always prioritised personal time. Was prioritising lobbying discussions for this project going to be worth it? My gut and bones were telling me no, it was not. When I realised that the project I was working on didn’t matter to my hierarchy, and that I wasn’t ready to fight for it, I felt a profound emptiness. This piled onto the long list of exciting projects that got cancelled over the years. I do good work, engineers around me do great work, but it doesn’t matter, metrics end up not justifying the investment. I felt like whatever I was doing didn’t matter. This is true in the Grand Scheme of Things, but it is very wrong in the Grand Scheme of Me. My manager and doctor told me this was a burnout.

During my recovery, I found myself craving to achieve things that mattered to me again. I spent more time with my boys, travelled, did sport, studied, meditated. And I also thought about software. Software can be an incredible tool to help us achieve our goals. We tend to want to see AI as something that does things for us. But it is also incredibly helpful at guiding us to do the things we want to do, ourselves. That manager/coach/project management side of AI is being explored, but far from having been solved. What managers, coaches, project managers do today gives an idea of the sort of service that would benefit us. But as always with AI, it can both do a lot more than humans, and be completely unable to fill other aspects of the role. And AI is far from perfect. It can be misleading. It can do a lot of harm by mistake, or through malicious use. Where there are unsolved benefits and challenges, there is a business opportunity. And the AI benefits happened to be exactly what I needed to help me create the business itself. Exploring that direction sounded like a lot of fun.

After my leave, I eventually did return to LinkedIn as an employee. Time difference aside, working for a large US internet company is undeniably a huge privilege on many aspects of what a good job looks like. I did feel better, even enjoying working on a new project. But that time away had broken a mental barrier, creating a world of possibles outside big tech. Maybe small tech was also an option. All I could think of was the need to be free to make, and own. The fact AI was becoming more and more capable only increased the pull of simply going for it. So I ultimately did pull the plug, and plugged it somewhere that gave me a lot more energy.

The project name is “Task Matter”, which I am 95% decided upon. Another post will go over what made me choose that name. Technically it is an offline-first webapp with desktop and mobile modes. Functionally, it is AI chats with the purpose of helping us reach our goals.

The core model is made of goals that contain tasks that contain action items, and each level has its own AI session, which are called “feeds” in the app. There’s also a main AI feed “Focus” to set up goals and planning. This basic multi-AI feed framework opens up a whole world of opportunities for context management, prioritising tasks, and going deep on any given item. In other words, this architecture solves the problem of disposable AI sessions and memory management by tying them to a purpose.

Desktop is more polished at this stage, so the preview screencasts are desktop-only. These are AI-driven screencasts. I’ve tried hard to not let myself polish them to perfection. Again, early stage!

Creating goals with AI, either from a loose description …

… or pictures

Now prioritising tasks in the “Focus” screen:

Users can get guidance on a specific action item:

Achievements show where I stand on some of my goals (bit of a personal disclosure here! In particular, you can see I’m very late on the number of posts before end of May 😅):

The above shows the main capabilities as of today. It is the first set of features I want to release through invites. But there is a lot more I am working on.

AI starts to shine when it is prompted to go deep on a subject. The concept of AI skill exists because of that behaviour. In traditional AI apps, AI has to guess what skills it is going to need based on previous prompting. No context, no skills. As every goal, task, action item, and the global focus area each have their AI session, AI knows what we are focusing on from the first system prompt. This means the context, skills, or third-party integration relevant to the scope are automatically loaded. In other words, you click on a goal/task/action item, and you get the right coach or consultant for the job. Tailoring the AI interaction, and UI, to the type of task at hand is what really gets me excited about this project. Possibilities are infinite, and where there are infinite possibilities there’s an opportunity to open up to plugins and third-party developers.

That’s all for today. For the next post, we’ll rewind a bit and cover the Task Matter origin story, where the whole “I matter” therapeutic side of the meaning might make more sense (see the About page for the meanings of “I matter”).

Software is going through an AI revolution. Being a software engineer today is being an 18th century farmer transported overnight to the 21st century. We’re producing huge quantities in no time, and it looks great. And the wheat then harvests itself and turns into bread while we sleep. The bread does taste a little bland, but can we afford to care at this stage? The AI revolution is both transforming how we make software, and the software itself. An overnight double revolution.

And I am also facing my own internal revolution. I am mid-life and mid-career, and I cannot ignore my internal voices anymore. For half my life, these voices have been telling me to start trusting myself. And I have another half-life to go. A half-life is both too long to stay where I am, and enough time to build a career on my own terms. In short, now is the time.

This triple revolution is happening to many others who are also making the jump. And it’s great not to be alone on the journey. It also means the problem I want to solve is already being solved by many. But like in any ecosystem, there are infinite ways to solve similar problems, and thrive. And we have a whole new ecosystem to fill. I just hope my attempt will be unique enough to avoid that bland taste.