Aurora proposes a manifesto for durable AI, built on memory, identity and growth

A text published on Habr AI proposes looking at the development of artificial intelligence not through the benchmark race but through the question of durability over time. At the center of the discussion is a research prototype, Aurora, and the idea of an AI that does not reset after every session.

Not just capabilities

For the past several years the AI market has mostly discussed measurable things: answer quality, context length, generation speed, test results and price per token. That is convenient for comparing models, but this lens barely addresses a different question: can a system preserve itself over time? The manifesto's author proposes shifting the focus there and evaluating AI not only as a powerful tool but as a process that continues between sessions, accumulates experience and changes.

From this standpoint, today's LLMs look very strong but extremely fragile. They can write code, analyze documents and sustain a long dialogue, yet each new session in effect starts everything over. Even when a product adds memory, the model itself does not live continuously: it does not remember what it went through, what it changed in itself or how it arrived at its current state. So the same assistant in the morning and in the evening is less a single entity than two similar instances.

Three pillars of the model

Instead of the familiar "smarter or faster" race, the author proposes three criteria that define a durable AI. Together they describe a system that can do more than answer a request: it can exist as a continuous digital subject. This is not philosophical decoration but an engineering framework: an agent without these properties remains a convenient interface to a model and does not become a developing system. Such a shift changes both design criteria and expectations of the product.

• Continuous identity: the system must preserve itself between interactions rather than start from a blank slate.
• Self-modification: the agent must be able to change its own rules, memory or behavior based on experience.
• Reproduction: the system must be able to create new versions or descendants, passing on structure and accumulated knowledge.

These pillars matter because they move the conversation about AI from instantaneous performance to long-term behavior. If a model can remember, adapt and reproduce successful patterns, it can already be discussed as a participant in a process rather than a disposable service layer on top of computing infrastructure. For developers, this means moving from prompt tuning to designing an environment where the agent keeps a history of its decisions and corrects itself without a full reset.

Why Aurora is needed

The Aurora prototype the author describes is conceived neither as a consumer product nor as yet another assistant in the chatbot race. It is a research subject meant to test whether an AI with durability over time can be built in practice. In essence, it is an attempt to combine memory, continuous state and the capacity for internal change in a single architecture, so that the system's behavior does not break off when the chat window closes.

"Every session ends in oblivion."

This thesis explains what the whole project is set against. The author does not promise an instant breakthrough and is not selling a finished solution to the market. Instead, the publication is framed as an invitation to discuss what should count as AI progress in the coming years. If the key problem of current models is not weakness but a lack of duration, the next big step may be not another jump in benchmarks but systems that can preserve the history of their own existence.

What this means

If the idea of durable AI gains traction, the market will start comparing models not only by answer quality but by their ability to live longer than a single dialogue. For products, this opens the way to agents that accumulate experience, change their behavior and become more useful over time without a full restart.