Rango: Why I Built a New Database for the Age of Agents
Rango: Why I Built a New Database for the Age of Agents
For weeks, I went down the rabbit hole. I was building stateful agents for Rumai Labs, trying to solve complex telemetry and decision-making problems in the field—places where internet is a luxury and reliability is a mandate. I needed a memory layer that was local-first, structurally aligned with AI reasoning, and durable enough to survive a process crash without losing its "train of thought."
I scoured the landscape. I looked at SQLite, MongoDB, and the current wave of Vector Databases. But after deep investigation, I realized that while these tools are masterpieces of engineering, they weren't designed for the hot path of an agent's conscience.
Nothing convinced me. So, I decided to build Rango.
The Search for the "Perfect" Memory
My requirements were simple but uncompromising:
- Zero-Latency hot path: Memory must be local-first. An agent shouldn't ask a cloud API in San Francisco for permission to remember what the user said two minutes ago.
- Structural Alignment: Agents think in BSON/JSON. Forcing that into relational rows (SQLite) felt like a hack that killed performance and flexibility.
- Durable History: I didn't just want the final state of a variable. I wanted the Oplog—the ability to "rewind the tape" and see how the agent evolved its reasoning.
SQLite is a legendary tool, but it's for relational data. MongoDB is powerful, but it's too heavy for the edge. Vector DBs are great for retrieval, but they are terrible for operational state. There was a void in the middle of the stack.
Forging the Solution
Rango is the answer to that void. It's an embedded document database written in Rust, designed specifically to be the circulatory system for stateful AI.
Why Rango is Different (The Numbers)
We ran a "fire test" comparing Rango against traditional Markdown-based memory and standard relational lookups for a common task: Retrieving a specific record from 1,000 logs.
| Feature | Markdown (Native) | Rango (v0.1) |
| Latency (ms) | ~2,500ms | ~7ms |
| Token Cost | High (Prompt Stuffing) | 0 (Local-First) |
| Structure | Unstructured Text | BSON Documents |
| Persistence | Fragile | Durable-First (Oplog) |
Beyond Data: Persisting Conscience
Building Rango wasn't just about speed; it was about sovereignty. In a world of centralized AI, the most valuable asset is the agent's memory. By making Rango BSON-native and UUID v7-driven, we ensured that memory is inherently chronological and easy to sync across nodes.
If SQLite is for storing data, Rango is for persisting conscience. It is the substrate that allows an agent to finally overcome digital amnesia and become truly autonomous.
By Antony Giomar - Staff Engineer at Rumai Labs.