The Beginning of Infinity

Context is all you need.

Introducing UltraContext, the context infrastructure for AI agents.

For the last 2 years, I caught myself reinventing the wheel for every agent I built. This REALLY annoyed me. Had an agent idea? Cool, before you even start building, time to reinvent a message storage relational model from scratch. Then, create a retrieval logic. Then add sessions logic on top of that. Then versioning. Then time-travel. Then the list goes on and on.

By the time you got done, the fun was over.

Why is this so broken?

Then you can make an argument on the using the existing frameworks. Yes, they exist. But they're trying to lock you in. The truth is everyone is desperate to capture the value of the AI era. Solving the problem? Nahh, that's secondary. Assuming they do, are you sure they will be around for the next 3 years? And assuming they will, would they build towards your specific needs? We both know the answer is no. The second you wanna go a little bit off the rails, they atrapalham more than help. Now you're locked in. And if your stuff turns out to be the next billion dollar agent, you'll have to rewrite everything to support a feature only your users really need. You're on the hands of the company that built the framework. How fun is that? We DONT need another agent framework. We need the right primitives. And they have to be agnostic. And WE should own it. They should be open source. They should belong to us, not to anyone else.

What impresses me even more is to see the entire industry building fancy abstractions like "memory" on top of a extremely broken, unreliable and non-standardized system instead of fixing the actual problem. You know what? The blinking neon graphs looks cool, but they don't fix my problem, and not anyone else's. The system still broken. Memory is just an example. There is dozens of other abstractions just like that, with hundreds of millions invested in them. You know why this happens? Because no one has a clue where this is going to, so they throw money in whatever impress them. We don't fully understand what we're living right now. Its unpcedecented, but everyones pretend its not. Those empty abstractions boosts valuations and engineers egos, but they dont fix the problem. When the bubble pops, most of these companies are going to zero. Mark my words. The truth is that engineers wanna feel they're doing something important. The truth is that AI now outworks even the smartests PHDs and folks are economically incentivized now to build complexity so they still valueble, because only they can fix the problem they ourselves created. Its very similar to what happens on Wall Street too. But the big ugly problem still there, and no one is solving it. I mean, it's 2026 and we still don't have a standard for context engineering. Let's just stop for a second to think how WILD is that. It's like we're still in the 90s, and for each new webpage we build, we have to recreate HTML from scratch. You cannot have react without html. Remember that.

It turns out that there is no standard for context engineering. So I decided to make it.

Our Goal: To make agents insanely sinple and pleasant to build by providing the BEST context infrastructure tools that makes you incredibly efficient so that you can ship agents for billions of users.

Every operation is automatically versioned with git-like semantics, but without the complexity. Branching, time-travel, rollback — all built in. Works with any LLM framework. TypeScript and Python SDKs from day one.

How it works

• create() creates a new context (or forks from an existing one with from).

• append() adds messages in order without creating extra versions.

• update() and delete() create immutable versions automatically.

• get() supports current state, full history (history: true), and time-travel with version, before, or at.

• Fork/clone from any point to test parallel agent flows without losing the original context.

SDKs

• JavaScript/TypeScript - npm install ultracontext (v1.0.1)
• Python - pip install ultracontext (v1.0.2)