For Marcin · Unabyss founder · Om Umrania
New conversation
// 9:07 AM · New Claude session · Third one this week Hi, I'm Om — I'm an AI engineer, I'm working on several projects right now, let me explain the context... // and so it begins. again. This is what every AI session looked like before Unabyss.
What changed
What happened when I called whoami

I stopped explaining myself
to my own tools.
That's the product you've built.

I'm Om — AI engineer, builder of personal AI systems, incoming MBA student. I used Unabyss as a Claude MCP server during a real working session. This is an account of that experience: what the problem is, how your product lands on it, where the depth is, and where I think it goes.

Why I started using it

I work across multiple active projects at any given time. I switch between them constantly. I use AI tools — not as occasional assistants but as the primary surface where I think, draft, and decide. Every meaningful working day involves multiple Claude sessions.

The problem is structural, not incidental. Every session starts at zero. Who I am, what I'm building, what I've decided, what I'm trying to accomplish this week — none of it carries. So every serious session opens with a tax: re-explain yourself, or accept a generic answer that doesn't account for who's asking.

"The bottleneck isn't the AI's ability to help. It's the AI's ignorance of who it's helping."

That framing is exactly what made me want to try Unabyss. You're not building a smarter AI. You're building the layer that makes any AI smarter about me. That's the right problem, and almost no one is working on it seriously.

How I actually use it

I don't use the web interface. I call Unabyss directly from Claude as an MCP server — meaning your tools are available inside my working sessions the way a file system or search would be. Here's what a real session looks like:

// session open — 9:11 AM whoami identity loaded · AI Engineer · Saathealth → Masters Union PGP list_integrations Google Calendar 3,473 events GitHub 2,139 files + pull requests Notion 145 pages + databases Gmail 85 threads agentic_query "what am I working on right now" synthesising across calendar + github + notion + gmail... answer ready · sources cited · context loaded query "what decisions have I made this week" quota_exceeded — Insufficient credit balance (preauth: 0.327)

Everything above the error line is the product working. The last line is the only moment it broke — and it broke in a specific way that matters: I had no idea I was spending credits, no visible balance, and no way to check before calling. When Unabyss is wired into a working session and a call silently fails, the session loses its thread with no recovery path. For someone using this as infrastructure rather than a demo, that failure mode is the most important thing to fix.

The reasoning behind using it

What I'm trying to build — in my work and in how I use AI — is a system that compounds over time. Not one where I repeat the same context-setup work every session, but one where what I've established, decided, and learned carries forward automatically. Unabyss is the first thing I've found that's genuinely pointed at that problem.

The synthesis that comes back from an agentic_query is meaningfully different from any single-source context. When it pulls from my calendar, my GitHub commits, my Notion databases, and my email threads simultaneously — it produces a picture of where I am that I couldn't assemble manually in under five minutes. That's the value. Not any individual source. The synthesis.

📅
3,473 calendar events — not a schedule, a behavioural map This tells you how I actually spend my time, what shapes my weeks, when I context-switch vs. go deep. That's richer than anything I'd tell you about myself.
💻
2,139 GitHub items — project memory, not just code Commits tell you where I've been. PRs tell you what I've decided and shipped. Having that retrievable means I don't re-explain my projects — they're already in context.
📗
145 Notion pages — structured intent Task trackers, project phases, planning databases. Whether you're indexing database properties (statuses, phases, dates) or just page titles determines whether I can ask "what's in progress" or only "what pages exist." That distinction matters a lot.
📝
The layer that's still missing — the reasoning Calendar tells you what I do. GitHub tells you what I build. Notion tells you how I plan. But the why — my decision principles, what I've ruled out and why, the frameworks I use to think — lives in local notes Unabyss can't reach yet. Context without reasoning is an activity log. With it, it becomes understanding.
The use case it's solving

I'm at a transition point — wrapping up a year of AI engineering work, starting a business programme in June, managing several long-horizon technical projects in parallel. The number of things I'm tracking is high. The number of AI sessions I run to manage all of it is also high.

Without persistent context, every session is isolated. I repeat myself, get answers that don't account for what I already know or have decided, and lose the thread between sessions. With persistent context, my AI sessions stop being conversations I start from scratch and start being continuations of something ongoing.

"I'm not looking for AI that can do more things. I'm looking for AI that already knows enough to do the right things — without me briefing it from zero every time I open a tab."

The concrete use case: I work on projects that span months. Coming back to something after two weeks means reconstructing context — re-reading notes, re-establishing where I left off, re-figuring out what the next step is. If Unabyss has been indexing my commits, calendar entries, and Notion updates across that gap, I can ask what happened and get a real answer. That's focus I don't lose. That's time that compounds instead of being spent on setup.

Where I think this goes

The version of Unabyss I'd describe in two years isn't a context vault you manage. It's a context layer that sits between you and every AI tool you use — so every tool always knows enough without you doing any setup work.

Already works
Session primer

One MCP call at session start. Who I am, what I'm working on, what the integrations know about my current state. This alone justifies the integration — it's a genuine shift from cold to warm in under a second.

Write-back from sessions into context

Right now context flows in one direction — integrations into Unabyss, Unabyss into sessions. But the decisions I make during a session, the directions I set, the things I figure out — none of that flows back in. Bidirectional accumulation is the difference between a snapshot and a living memory.

Local reasoning layer — even a CLI push is enough

Not asking for a full vault sync. A targeted push of specific decision logs and principles would close the biggest remaining gap. Everything else is activity. This is understanding. And it's where the competitive moat is — no one else is indexing the local reasoning layer.

The real opportunity
Context as infrastructure across all AI tools

Right now Unabyss is one tool among many. The version that becomes irreplaceable is one where every AI tool I use — not just Claude — pulls from the same context layer. Consistent context across sessions, across tools, across time. That's not a context vault. That's a new category.


Om Umrania
AI Engineer · Saathealth
Incoming PGP TBM · Masters Union, June 2026
omumrania2020@gmail.com
Happy to continue this — either as a follow-up to whatever you were looking for from this feedback, or as a deeper session on how the MCP integration behaves inside a real working setup. I've been using Unabyss as infrastructure, not as a demo. The findings above come from that.