v0.1.2 · Open source · Very early dev
Instead of asking one big AI model to do everything, Nanana runs a set of small focused models, each doing one job well, on top of a shared layer that handles memory, routing, and logging. The whole thing runs on your own hardware. No cloud. No subscription.
SoS stands for School of Specialists. the architecture principle behind Nanana. Each specialist does one job. The infrastructure ties them together.
One big model handles everything: memory, reasoning, tool use, personality, all crammed into a single context window. When something goes wrong, it is hard to tell why.
Build the infrastructure first (memory, routing, logging), then drop in small focused models as tenants. Each one does one job. When something breaks, you can see exactly where and why.
The key finding from production testing: a failure that looked like the AI reasoning badly turned out to be a small retrieval component returning the wrong input. Every model was behaving correctly. The problem was visible and fixable because the pieces were separate.
That is what transparency buys you. Not just for debugging, but for trusting what your local AI is actually doing.
runs on any machine with Ollama. Bring your own model. use what your hardware can handle. The infrastructure is model-agnostic by design.
A standalone shadow logger that sits alongside any AI and records every interaction as plain text files you can read yourself. Useful for understanding what your AI is doing, auditing past sessions, and carrying memory across different models without losing anything. Built for use outside the SoS infrastructure too.
Drop it alongside whatever you are already running. Logs input, output, and tool usage as an outside observer.
Inside the SoS infrastructure, it sees everything: context assembly, routing decisions, each specialist's input and output.
Nanana is a personal research project, now open source at v0.1.2.
It runs on consumer hardware. Built on a decade old PC because
nothing else ran on it. Has produced findings worth sharing.
Two papers are published documenting the architecture and what
production use revealed.
It works. It is not polished. The infrastructure is stable enough
to run daily, rough enough that you should expect to read the
source when something breaks. The papers are the documentation.