PersonalOS v2: Giving AI more room to work

PersonalOS is entering its next phase. The system I started building to preserve context between AI conversations is becoming something I want to rely on throughout the day: remembering what matters, helping me decide what deserves attention, and taking care of more of the work around the work.

The biggest change is how much authority I’m willing to give it.

In my first build-log entry, I wrote that “AI proposes, humans decide,” and meant approval for every change. I wanted the system to understand my decisions, but I was cautious about letting it organize or update that understanding on its own.

After spending more time working with different models and AI tools, I’ve come to trust their judgment with my guidance. The restrictions that felt appropriate earlier were increasingly slowing me down. I was reviewing routine work that I was comfortable letting the system handle.

So the principle behind v2 has changed: AI works autonomously within guardrails. I set the direction, review exceptions, and stay involved where judgment matters most.

That trust comes from using it. One of the biggest wins so far is that I can plan, deploy, and launch without being at my primary computer. I can move work forward from another device instead of waiting until I’m back at my desk.

The next step is to bring that same continuity to more of my day. That requires some changes underneath it.

The foundation remains a vault of plain-text files, with Git keeping a history of changes. This is where durable knowledge belongs: project context, preferences, decisions, and the reasons behind them. Obsidian gives me a way to browse and navigate it. The files remain readable outside any particular application.

That matters because I expect to keep changing models and trying new AI tools. My understanding of a project should survive those changes. A new assistant should be able to read the context and get oriented without requiring me to reconstruct months of conversations.

MCP, short for Model Context Protocol, is part of how I’m designing that access. It gives compatible AI applications a standard way to call tools and retrieve information from another system. In this architecture, it provides a way for an assistant to ask the vault for relevant context through a defined interface.

For me, the practical benefit is continuity. I want to ask a question from whichever AI tool or device I’m using and have it draw on the same underlying information. MCP helps make that possible, while the system still has to define what each tool can read or change. Connecting an assistant does not, by itself, give it permission to do everything.

The hardware has distinct jobs too. My MacBook is becoming my primary working machine. The Mac mini is the always-on home for background processing. My phone and iPad are places where I can ask questions and move work forward. The v2 design separates those roles so that routine processing can continue when I close my laptop.

I’m also being more deliberate about which system owns which information. Meeting tools such as Granola capture conversations. Todoist is the planned home for actionable tasks. The vault holds the lasting context that helps explain those tasks and decisions. Each should do its job without requiring me to maintain competing versions of the same to-do list.

Performance is another reason for v2. As I used PersonalOS more, v1 was starting to feel slow. Having the context available was useful, but waiting for the system to find it and assemble an answer was becoming friction of its own. During a meeting, an answer needs to arrive while the conversation is still relevant.

That changes how I’m approaching retrieval. The v2 design moves more preparation into the background, keeping searchable indexes current as information changes and giving the assistant a focused path to the context it needs. The goal is to spend less time rediscovering where things live each time I ask a question.

The plan combines exact searches for names and phrases with searches based on meaning, so I can retrieve a decision even when I don’t remember the words used to describe it. Those indexes can be rebuilt from the underlying records. The records remain the durable foundation.

I’m also treating speed and answer quality as things to measure together. A fast answer that misses an important decision is still a failure. The test is whether the system can find the right context, show where it came from, and respond quickly enough that reaching for it becomes a natural part of my day.

All of those choices serve a fairly ordinary moment: sitting in a meeting and needing to remember something.

I want to be more engaged because I’m listening, thinking, and asking questions instead of trying to capture everything. The information should still be there afterward, along with the decisions and follow-ups. During the conversation, I want to ask my vault what we previously discussed, why we made a decision, or what remains unresolved, and get an answer quickly enough to be useful.

The same context should make everyday administration easier. Scheduling, routine email, messages, and follow-ups all take more effort when I have to reconstruct the background first. I want PersonalOS to bring that background forward, guide my responses, and handle routine steps within the authority I’ve given it.

My to-do list should benefit from that understanding too. A task’s importance depends on what has changed, who is waiting, and what I’m trying to accomplish. I want help prioritizing and deprioritizing as those conditions change, without having to explain my entire week again.

When I wake up, I should have a clear view of what is ahead of me. As I move from task to task, I want to focus deeply on the thing in front of me, knowing I can recover the context for the next one when I need it.

There is a similar opportunity in my network. I want a clearer picture of who I’ve been talking with, what those conversations were about, and who I should reconnect with. When a relevant problem comes up, the system should help me remember the person whose experience or perspective could be useful.

My rental properties and future investments are another practical use. I want the vault to carry the context of the portfolio, the outstanding tasks, and the criteria I use to evaluate a purchase. That would let me resume a property analysis without rebuilding my assumptions, compare an opportunity against what I’m actually looking for, and surface questions I might otherwise miss.

Giving the system more authority also changes what I need to see. The architecture calls for a controlled process for writing changes, a record of what happened, and a dashboard on ProjectWong that shows whether background work is running or needs attention. Change history and backups support recovery. Those mechanisms are what make greater autonomy practical.

V2 is still being built. The cross-device productivity is already real. The connected daily experience, the retrieval improvements, and the expanded automation are the direction of the current work.

The measure of success will be how it feels to use on an ordinary day: whether I’m more present in a meeting, whether a follow-up happens, whether I can resume a project without retracing my steps, and whether I spend less time maintaining the system.

I’ve developed enough trust to give it more responsibility. Now I’m building the version that can put that trust to work.