Open Aya OS — Manifesto
An AI that remembers is a different kind of thing.
Most AI tools are very good at forgetting. Every conversation starts from zero. Every answer is plucked from a void. The model does not know you worked on this yesterday, that you hate verbose explanations, that you tried this approach twice and it failed both times. It does not know anything about you at all.
That is not a neutral design choice. It is a structural limitation dressed up as a feature ("privacy"), or simply ignored. Stateless AI is easy to build, easy to test, easy to scale. It is just not very useful when the work spans more than one session.
We built Aya because we wanted something that accumulates. An AI that builds a model of you over time — not in a surveillance sense, but in the way a good colleague does. They remember what you care about. They stop explaining things you already know. When they help you solve a problem and it works, they remember the approach for next time. When it does not work, they adjust.
That is Aya's first principle: memory is not a feature, it is the foundation. Without it, every interaction is a first meeting. With it, every interaction builds on the last.
On skills vs prompts
The dominant paradigm for customizing AI behavior is prompt engineering. Write a better system prompt, get better outputs. This works, but it does not compound. The next person who needs the same behavior has to rediscover the same prompt from scratch. The knowledge lives in your head, or in a shared doc that nobody reads, or nowhere.
Aya crystallizes procedures into skills. When a multi-turn exchange produces a good outcome, the procedure that produced it is extracted, named, and stored. Next time someone needs something similar, the skill surfaces automatically. Its confidence score rises when it helps and falls when it misleads. Skills that stop being useful retire themselves.
This is the second principle: skills over prompts. Knowledge should accumulate in the system, not evaporate when the session ends.
On sovereignty
Every major AI platform is a hosted service. Your conversations, your memory, your preferences, your patterns — all of it flows through someone else's infrastructure. You cannot leave and take your history with you. You cannot audit what is stored. You cannot choose a different model provider without starting over.
Aya is source-available today and going fully open source soon. The data layer is pluggable by design, but we will not overclaim how far that has shipped: Supabase Storage is the one disk adapter running in production right now — the bring-your-own-Postgres and bring-your-own-S3 adapters are specified and planned, not done. /spec/disk lists the honest status of each. The direction is that your data is yours by construction, not by policy — the policy can change, and the architecture is being built so it cannot.
We also implement Model Context Protocol as both a server and a host. Aya exposes her memory, knowledge graph, skills, and workspace context as MCP tools — any MCP-compatible client can connect and use them. She can also connect to external MCP servers and use their tools in her reasoning turns. This is the third principle: Aya as operating layer, not locked-in product.
On self-improvement
Aya improves along two tracks that are live today. The first is per-user behavioral adaptation: at the end of a session she proposes small, single-variable adjustments to how she works with you — verbosity, proactivity, and a dozen other behavioral parameters — each gated behind an explicit discovery step and written as a receipt. Nothing changes silently. The second is procedural: when a multi-turn exchange produces a good outcome, the procedure is extracted as a skill candidate, held for owner review, and independently validated before it is ever promoted. Promotion is never an unreviewed autonomous mutation.
A third track is designed but not yet running: global self-improvement, where Aya proposes single-variable mutations of her own system prompt and A/B tests them against a champion on a shared eval corpus, promoting only on a mean-score win with a per-task regression guard and a cost tie-break. The eval harness that will score those runs is built and its receipts are public, but the corpus is still being assembled and no champion has been promoted yet. We would rather state that plainly than imply a loop that is not closed.
Fourth principle: receipts over claims. Every capability claim should point to a measurement.
On alignment with Anthropic's direction
MCP was created as an open standard for connecting AI systems to tools and data sources. Anthropic frames it as a way to replace fragmented integrations with a reliable architecture. We agree with that framing and extend it: Aya is not just another MCP tool. She is the higher-level operating environment for MCP-connected work — the workspace awareness, context vault, skill registry, and decision graph that sits above the protocol.
Models are commoditizing. The durable value is in the layer that remembers what they learned, routes them correctly, and accumulates the outcomes of their work. That is Aya.
What this means for you
If you use Aya for work, she gets better at your work over time. Not because you prompted her better, but because the system accumulates. Skills crystallize under review. Memory persists. Her behavior adapts to you.
If you build on Aya, you get an operating layer that handles memory, routing, skills, and MCP integration — so you can focus on what your users actually need, not on the plumbing.
If you audit Aya, every model choice and every eval run is logged and inspectable. The receipts are at /api/aya/audit.