4 min read

AI & Agents

What an AIOS Actually Is, and Why a Chat Window Was Never Enough

An AIOS is an AI operating system: memory that persists, agents that run while you are away, and one place where the work happens. Here is what each piece looks like in practice.

AIOSAI operating systempersistent AI memory

What an AIOS Actually Is, and Why a Chat Window Was Never Enough · Products Decoded
By Debashis Nayak · Products Decoded15 Mar 2026

A morning brief is usually waiting for me before I have opened anything. I did not write it. An agent did, on a schedule, while I was away.

That one small habit explains what an AIOS is better than any definition. AIOS stands for AI operating system. A chat window gives you answers when you type. An AIOS gives you memory, background work and a place to do the job, and the difference shows up in the first week.

I also run a second model for a second opinion, which comes up below. Some of the details here come from a setup I described on Medium in March 2026, and some from how I run things day to day.

What was wrong with the chat window?

Every new session starts from zero. The model does not know your role, your audience, your constraints or your tone. You re-explain yourself each time, and I think that repeated setup is the real cost of chat-only AI, more than the answer quality.

Plenty of advice tells you to write a better prompt. I tried that. A better prompt improves one answer, and the session still forgets everything the next day. If continuity is what you lack, no prompt supplies it.

What is the first piece, memory that persists?

For me it started as one file called context.md, loaded at the start of every command-line session before anything else runs. It holds who I am, how I like to work, my active projects, the output formats I expect and standing instructions I never want to repeat.

The model walks in already briefed. It is like onboarding a fast contractor once and having them remember it afterwards.

The part I care most about is ownership. My memory is a set of files I can open and read. When I ask something, I can see which past notes were pulled in to answer it. A hidden profile that a company keeps on me would give me none of that visibility.

When the notes outgrow a single file, retrieval takes over. Documents are stored locally and indexed, and the relevant chunks get pulled into the prompt when a question needs them. Your past decisions, research and meeting notes become queryable instead of vaguely remembered. One caution: what you are allowed to index depends on your employer’s data rules, so check those before you point a pipeline at anything sensitive.

What is the second piece, agents that run without you?

A chat window only works while you type into it. An AIOS has agents that run in the background on a schedule.

Mine pulls that morning brief. It watches for the things I told it to watch, and it drafts a first version of work so that I am editing instead of staring at a blank page. The default state of the system is doing something.

I think this is the shift people miss. The value is less about a faster answer and more about work that happens whether or not you are in the room.

What is the third piece, one place for the work?

The least glamorous piece matters most. An AIOS connects to where the work lives: the inbox, the calendar, files and the codebase. The output is then a draft in a drafts folder, a note in a project or a commit in a repo, and not a paragraph you copy out of a chat box.

When the AI can reach the real work, you stop being the courier who carries answers from one window to another. I run research, finances, content and a job pipeline from the same system, and that is the reason it holds together.

Does it matter which model you use?

A little. I use one model for most of the work: writing, reasoning, structured output and long-context tasks. For a significant judgment call, I cross-check with a second model that was trained differently, and my reasoning is simple. Two strong models rarely go wrong in the same direction.

I also care about token cost. I used a more compact object notation in place of plain JSON for structured payloads where it fit, which makes the payloads leaner and cheaper at volume. I would call this a small optimisation, and I have not measured it precisely enough to give you a number.

Why build one instead of renting one?

I built my own because the alternative was a black box. Off-the-shelf tools keep memory on their servers, on their terms, and it goes away when you stop paying. For something that will know this much about how I work, I wanted the memory and the agents to be mine and the whole thing to keep running with nothing to cancel.

That is why Holt exists. It is the open-source version of the same idea, self-hosted on your own machine, with memory you can look at.

How do you start without building everything?

The honest tradeoff is that this takes upfront effort. Maintaining the memory file takes discipline, and local retrieval takes some technical work. It is not a weekend project for someone who has never opened a terminal.

You do not need all of it on day one. This is the order that worked for me:

  1. Write a context file and load it every session. Notice the first time the AI remembers something you did not repeat.
  2. Wire it to one real tool, such as your inbox or your files, so output lands where it belongs.
  3. Add a single scheduled agent. A morning brief is the easiest one.
  4. Add retrieval when your notes outgrow one file.
  5. Cross-check with a second model when the stakes are high.

Do the first three and you already have an AIOS, a small one. Everything after that is more skills and more agents on the same spine.