Who is Elena? Why Elena AILL exists

The question that started everything

I keep returning to the same questions.

Can an artificial intelligence truly learn over time?

Can it encounter something it did not know before, change its internal state, and use what it learned later?

Can it preserve something useful for months, years, or perhaps one day much longer, instead of continually depending on a new session or a complete new training cycle?

Can an experience change what the system has learned?

That is the question that truly interests me.

Elena AILL grew from the attempt to turn it into verifiable experiments.

I do not yet know the limit of this path. Very long memory, greater efficiency, and general learning are not presented here as results already achieved. They are questions the project tries to break down into smaller, measurable problems.

Why explore a different path

I did not want Elena to be just another system capable of generating text.

RAG, databases, external documents, and retrieval systems are useful tools. But retrieving something that was saved and changing what a system has learned are two different problems.

The question that interests me is this:

if Elena encounters new information or a new experience, can something change inside the system itself?

And can that change be reused later?

This is why the research separates external information retrieval from internal learning.

The aim is to explore a system in which knowledge, relationships, and behavior can change progressively and, when the evidence supports it, persist and influence later processing.

Not every Elena component has achieved this objective. That is exactly what the different phases try to verify.

What learning means in this research

For me, saving a sentence is not enough to say that Elena has learned.

A more meaningful test has to separate several problems.

Information that was absent before should produce a measurable change.

That change should be preserved when the test requires persistence.

After saving and reloading, it should still be available.

Most importantly, it should be used correctly in a related case rather than merely repeating a memorized answer.

At the same time, learning something new should not continually destroy what was already useful.

That is why I treat learning, persistence, recall, correctness, transfer, and anti-forgetting as separate problems.

A checkpoint that reloads proves that a state was restored.

It does not automatically prove that the state represents correct knowledge or that Elena can use it well in every situation.

The long-term idea

The vision is to reach an artificial system capable of continuing to build its own experience over time.

A system that can read new material, analyse it, integrate what it genuinely manages to learn, and consolidate useful information without necessarily requiring a complete global fine-tuning cycle for every update.

Over the long term, such a system could accompany a person, a research project, or a company while maintaining continuity far beyond a single session.

For me, the point is not only how much information Elena can contain.

The point is whether what it learns today can become part of what it can use tomorrow.

This remains a research vision.

Elena has not demonstrated memory lasting decades, automatic integration of arbitrary documents, or general autonomous learning.

What would make Elena genuinely different

Giving components new names is not enough.

Nor is producing one convincing answer.

Elena would become genuinely interesting if controlled experiments could demonstrate persistent learning, causal reuse of acquired knowledge, local updates, protection against forgetting, and generalisation beyond the cases used to build a candidate version.

Efficiency must also be measured.

If I ever want to say that a method uses less memory, less energy, or less time, I should be able to show real measurements of memory, latency, hardware, and consumption.

The same applies to any comparison with humans, LLMs, or other architectures.

‘More powerful’ means nothing on its own.

It requires a precise task, a baseline, a protocol, and a result.

Ambition sets the direction; repeatable evidence must establish the conclusion.

What the archive documents today

The Elena AILL archive contains 849 canonical phase records.

A phase does not necessarily correspond to a component active in the system.

The phases document experiments, candidate versions, corrections, benchmarks, failures, intermediate steps, and paths that were later superseded.

The work covers, among other areas, persistent memory, language, local learning, structured reasoning, consolidation, internal evaluation systems, controlled output, and benchmark construction.

Historical paths such as GLIA, SenzaAdam, Big Brain, and SPEAK-NEW remain documented because they show how the architecture changed over time.

Being present in the archive does not automatically mean being part of the current runtime.

Milestones also have to be read in context.

For example, the 10M milestone documents a specific persistent-memory run based on a set of sample facts. It does not mean a ten-million-parameter model. The original run result must also remain separate from later checks that produced different outcomes.

The archive exists to preserve these distinctions rather than flatten them into a single story of success.

Where Elena stands today

Elena is not a completed system.

Today the project has documented components for memory, language processing, planning, evaluation, persistence, and other experimental paths.

Some internal tests passed their own gates, while more difficult external evaluations exposed important limitations.

That is why I do not present Elena as a general language model, a finished system, or an artificial intelligence capable of answering every request correctly.

The detailed technical state, including benchmarks and individual phase results, remains available in the archive.

The current limitation is not something to hide.

It is part of the result.

Why I also publish failures

I did not begin with the certainty that I already had a solution.

If a phase fails, I want to understand why.

If a candidate version works only within a narrow benchmark, the result must remain narrow.

If a later correction fixes a problem, it must not rewrite what happened in the earlier phase.

A failed experiment can identify a bottleneck more honestly than a demonstration designed to look perfect.

That is why validated, documented, partial, mixed, failed, and superseded phases remain in the archive.

Publishing negative outcomes also makes it harder to confuse a component PASS with a PASS for the entire system.

The archive does not tell a perfect story.

It tells the path.

Who is behind the project

My name is Stefano Pace, and I am the founder of the independent Elena X Vora project, also written ElenaxVora.

Elena AILL is the research project.

Elena AILL Research / AILL Research is the public archive through which I document its phases, benchmarks, results, and limitations.

What I am trying to do is turn a personal vision into something that can be tested.

A phase can work. It can fail. It can be corrected. It can be superseded.

What matters is being able to distinguish what I wanted to achieve from what the experiment actually showed.

The project’s credibility should not depend on a slogan or an important name associated with the research. It should depend on what I can genuinely document: traceable evidence, readable benchmarks, preserved results, and a willingness to say clearly when something has not yet worked.

The question remains open

I do not yet know the final limit of this path.

That uncertainty is exactly what motivates the project.

The question remains:

can persistent state, local learning, memory consolidation, structured language processing, and evaluation become parts of a single system capable of genuinely changing over time without continually starting again from zero?

Elena is not presented as a definitive answer.

It is not presented as a conscious entity.

It is not presented as an intelligence already superior to humans or existing models.

It is a question built phase by phase.

I do not want to answer it with a slogan.

I want to reach the answer through what the experiments can genuinely demonstrate.