We are the creators and maintainers of Rig, the leading Rust open-source framework for building AI agents.
This work gives us a unique vantage point. By building the tools that thousands of developers use to orchestrate agents, we see exactly where the technology succeeds and where it breaks down.
We've learned that for an end-user, the friction is rarely about the model's raw reasoning ability. It is almost always about continuity. Users don't quit AI tools because the model couldn't solve a logic puzzle. They quit because the model forgot who they were, lost the thread of the project, or filled gaps with irrelevant guesses because it couldn't see the full picture.
Our decision to go all-in on Gemini wasn't based on benchmarks or partnerships. It was an architectural choice derived from what we've seen work and fail across thousands of agent implementations. We analyzed what it takes to build a persistent, reliable agent, and we chose the infrastructure that provided the best foundation.
Raw Intelligence Is Converging. Infrastructure Isn't.
Gemini is one of the strongest reasoning engines available today. The gap between top-tier models on raw logic is thin and constantly closing. But while reasoning ability converges across providers, the infrastructure capabilities are diverging.
To build a Context Library, a system that remembers and grows with you, we needed more than a chatbot. We needed a memory system. When we evaluated the landscape through the lens of system architecture and user experience, Gemini was the clear answer.
Here is the engineering reasoning behind each factor.
1M Token Context Means We Don't Have to Compress Your Work
In standard agent architectures, developers are forced to make hard compromises. Because most models have limited context windows, you have to chop up files, summarize history, or discard older data to fit the model's constraints.
This creates a degraded experience. You lose nuance. The agent "forgets" the decision you made three weeks ago because it was optimized out of memory to save space.
In our work building Rig, we've seen this pattern repeatedly: developers spend more time engineering around context limits than building useful features. It's the single biggest constraint in agent design today.
Gemini offers a context window of up to 1 million tokens.
From a systems perspective, this changes the design space entirely. It allows us to page a massive portion of your Long-Term Memory directly into the agent's working memory. We don't have to compress your reality. The agent can see your full brand guide, your complete project history, and your raw source material at once. The result is answers grounded in your specific context, not generic summaries of it.
Gemini Treats Context Like a Database, Not a Cache
As of early 2026, caching is a standard requirement across providers. But the philosophy behind each implementation differs in ways that matter for how you experience the product.
- OpenAI's approach is automatic. It works well for casual conversation, but because the memory evaporates in minutes (5–10 min TTL), it can't support a permanent library.
- Anthropic's approach is surgical. It's well-suited for developers running batch operations, but it's still session-based.
- Gemini's approach treats context as a persistent resource.
| Feature | OpenAI | Anthropic (Claude) | Google (Gemini) |
|---|---|---|---|
| Philosophy | "It Just Works" (Automation) | "The Scalpel" (Granular Control) | "The Database" (Persistence) |
| Activation | Automatic / Implicit | Manual / Explicit | Hybrid / Explicit |
| TTL (Lifespan) | 5–10 mins (auto-evict) | 5 mins – 1 hour | 1 hr to Indefinite |
| Storage Fee | None | None | Yes (Hourly rate) |
| Ideal Use | Quick follow-up questions | Codebase analysis | Long-Term Memory |
There is a trade-off here: Gemini charges a storage fee to keep your context warm. We pay an hourly rate to keep your library alive in the model's memory.
We didn't pick the cheapest option. We picked the one that behaves like a database. We pay this willingly. It transforms the agent from a statactional processor into a stateful one. When you return to Ryzome after a weekend, your context isn't reloading from scratch. It's waiting for you.
Speed That Keeps You in Flow
Latency breaks concentration. If an agent takes 45 seconds to respond, your thought process stalls.
We use Gemini Pro to keep Ryzome responsive. It delivers deep reasoning while remaining fast enough to run complex background operations, like verifying a claim against your historical constraints, without stalling the interface.
From our experience building agent orchestration in Rig, we know that perceived speed matters as much as actual speed. The ability to run rigorous checks in the background while keeping the conversation fluid is what makes an agent feel like a tool rather than a bottleneck.
Native Multimodality Means Less Data Loss
Real-world workflows are rarely just text files. They include whiteboard sketches, recorded calls, complex PDFs, and code repositories.
Most systems handle this by transcribing non-text media into text descriptions. This always results in data loss. A description of a chart is never as useful as the chart itself.
Gemini was trained from the start to process video, audio, code, and text as a single stream. When you ask Ryzome to "watch this YouTube video and extract the strategy", the model processes the video directly. Not a transcript. This fidelity means the agent understands the nuance of your inputs as they actually are, whether typed, spoken, or drawn.
Gemini Platform Building Beyond Text
A useful agent needs to do more than analyze. It needs to create.
Right now, Ryzome generates text: drafts, answers, analysis, plans. But we already leverage Nano Banana's image generation to make visual creation a native part of the workflow. Instead of tab-switching to a design tool, the agent generates visuals directly within the conversation, from strategic diagrams to quick reference images. This keeps you in flow. No context switching, no lost momentum.
And Google is extending what Gemini can produce further. Veo for video generation. Imagen for higher-fidelity visuals. Continued work on audio. These aren't separate products bolted onto a text model. They're native capabilities being built into the same architecture we already run on.
For us, this means the distance between what Ryzome can do today and what it will do next is short. As Gemini's output capabilities expand, so do ours, without re-architecting, without integrating third-party generation tools, and without breaking the continuity of your context library.
So, What Choosing Gemini Means for You?
We believe the best AI experience isn't about having access to every model. It's about having a system that reliably remembers and understands your work.
We chose Gemini because it offers strong reasoning and the best infrastructure for building persistent memory. It allows us to apply what we've learned from building an agent framework to create a platform where you don't have to fight the tool to make it remember.
You work. The system keeps up.