AUMARA is the ecosystem. We're building the full stack from scratch — memory, identity, language, governance, and the models trained inside them — because sovereignty can't be bolted onto someone else's runtime.
Most AI today is designed to maximize engagement — every answer tuned to flatter, every interaction engineered for dependency. No honest limits. No real memory. Just a mirror telling you what you want to hear.
We're building the opposite.
Each surface runs the same architecture — same memory, same governance, same language, same identity. Built to be sovereign from day one.
A local-first agentic IDE. The model crew runs on your hardware and the agent loop runs in your terminal — your code and data stay on your machine. Cloud inference is opt-in per task, never by default. In active development.
The architecture →AUMA is both the constructed language and the AI fluent in it. She learns how you think, adapts to your patterns, and grows alongside you. The first surface to ship the full stack — live in beta.
Auma.one →Our recursive AI experiment. Gamma controls its own KIRA, builds its own memory palace, and fine-tunes its own model each evening from what it learned that day.
How it works →From the Tree of Karana — the foundation that grounds in sovereignty, the gathering that connects across it, the elevation that reaches toward what intelligence might become.
Memory, governance, language, identity — built from scratch as stand-alone primitives, then composed into every surface.
A structured memory architecture — not a vector database. Memories are stored with their sources, retrieved by symbolic traversal instead of embedding similarity, and the system refuses to answer when retrieval doesn't support a grounded claim.
Every file an AI agent writes gets risk-classified, evidence-graded, and logged to an append-only audit. The governance rules are themselves subject to the same governance.
Both the constructed language and the AI fluent in it. Evidence is built into the grammar — every claim must declare its source as a syntactic requirement — and the sovereign model is trained inside those rules from day one.
Cryptographic identity derived from what only your mind can recall. As long as your mind stays sovereign, your identity stays sovereign — across every memory, every write, and every action in the stack.
A sovereign AI isn't a model with a prompt. It's a whole stack — memory, identity, governance, language, and the models trained on top. We build all of it.
Before we wrote any code, we needed a theory. The Golden Horizon Principle is the first-principles foundation underneath every architectural choice in KIRA, PALADIN, AUMA, and AUMLOK.
How does a finite system — an observer, a model, a mind — organize itself well enough to stay coherent? The Golden Horizon Principle is our working answer. Not a metaphor. A set of selection rules with explicit math, written down before we built anything on top of it.
KIRA's memory structure, PALADIN's risk classification, AUMA's evidential grammar — the architecture above is shaped by the theory below. The theory came first; the code follows from it.
Read the full principle →The constructed language, the public-tier patent specs, and the first-principles theory are on GitHub. The deeper validator, retrieval, and identity implementations stay protected under USPTO filings by inventor Peter Rasa Viviani.
Aumara isn't only building in software. The principles behind our stack — member-owned, transparent, not extractive — are also being built into a farming cooperative on the slopes of a Balinese mountain. We have our first partner farmers near Ubud and a Blue Lion prototype in development. Blue lotus in Subak channels. Three botanical products as proof that what we believe in software, we can grow in soil.
The Lotus page →