What does it mean for an AI system to learn? Keeping a piece of information is a start. The system also needs to find it when it matters and put it to use.
That question is central to Elena AILL, an independent research project investigating memory, language and learning through documented experiments. Two experiments make the importance of those separate steps especially clear.
In F898, a correction takes the checks for available linguistic forms from 14/28 to 28/28. The meanings those forms need to express have already been structured.
In F896, the planner passes all 32 checks when it receives organised information. Starting with requests in natural language, however, the complete path runs into much greater difficulty.
These results show why improving one function does not necessarily improve the system as a whole.
01 / THE ORIGINWhat should an experience leave behind?
The public project introduction concerns a change that remains useful: a system encounters information, retains something from it and can use it later. The revealing moment comes after that first encounter, when the question is worded differently or the information becomes relevant in a new context.
An archive can contain the right fact without the system being able to find it. Retrieval can succeed without producing a relevant answer. Elena AILL tackles these difficulties by separating the steps and testing them individually.
The project archive follows that work, including candidate versions and experiments that did not pass evaluation. It turns the broad question of learning into specific problems that can be investigated.
02 / THE DISTINCTIONSKeeping, finding and learning
Memory concerns what remains available. Retrieval concerns how it is found. Learning raises a further question about behaviour: after the experience, how does the system handle the task being studied?
Suppose a fact is stored correctly but retrieved to answer the wrong question. Retention has worked; selection or interpretation has failed. Calling all of this “learning” would hide the distinction needed to investigate the error.
A change of state needs the same care. Writing a record or adjusting a priority establishes that something was updated. Finding out whether it helped requires a comparison of behaviour before and after, under the conditions of the test.
What happens over time matters too. Does the improvement survive a reload? Does it disrupt earlier knowledge? Where experiments include these checks, persistence and preservation of previous behaviour are part of the finding.
The pages on persistent memory, continual learning and internal update mechanisms explore these different sides of the problem.
03 / A POSITIVE RESULTF898: supplying the missing forms
In Phase F898, the meaning is already organised. The task is to make the linguistic forms needed to express it available, preserving the required relations and arguments. This is the point in the chain that the test examines.
The sample stays fixed at 28 structured meanings. The reference meets the availability check in 14/28 cases; the F898 candidate reaches 28/28. The correction covers every case in that check.
F898 · Available linguistic forms
Reference: 14/28. F898 candidate: 28/28. The same 28 structured meanings.
The chart adapts to your screen. You can also .
This is a specific improvement whose location matters. Final answers to open-ended questions are not evaluated for these twenty-eight cases: the test checks whether the required forms are available. Understanding a request and building its meaning happen earlier.
The phase records PASS, while the lexical candidate remains unpromoted. A successful experiment and adoption in the operational version are separate decisions.
04 / AN INSUFFICIENT RESULTF896: why a good plan is not enough
Phase F896 examines two starting points. In one, the planner receives the problem and evidence in structured form. In the other, the system starts with a written request and must build that representation before it can reach an answer.
With the prepared input, the planner passes 32/32 checks. That result describes its work with evidence supplied by the test; it does not include the interpretation already performed upstream.
F896 · A component is not the whole system
Structured planner: 32/32. Complete interpretation: 3/32. End-to-end responses: 2/28. Distinct checks and denominators.
The chart adapts to your screen. You can also .
Starting from text, interpretation is complete in 3/32 cases. End-to-end answers succeed in 2/28 applicable cases, excluding those for which the test calls for deferring an answer. These fractions describe different checks. The overall gate remains unmet, with an INSUFFICIENT outcome and no operational promotion.
F898 addresses a later gap in the availability of linguistic forms. Its improvement does not change F896’s historical verdict: finding the words and understanding the request need different tests.
05 / THE CONTEXTFrom a research record to a working system
Following research also means keeping track of versions. A candidate may work in an experiment without being adopted; a component may exist in the project without playing the same role in the operational path.
F896 and F898 are historical experiments. The architecture page provides the accompanying account of components and runtime boundaries. These results alone do not establish what the current chat can do.
The language system hub traces the links between interpreting a request, organising information and producing a sentence. For an answer to be useful, interpretation, information organisation and language production need to work together.
06 / THE CONCLUSIONFollow the change all the way to the answer
Fluency is immediately visible. A useful change takes closer examination: which error has been corrected, what happens under different conditions and what remains of previous capabilities?
F898 shows a gap being closed. F896 shows the distance between an effective component and the complete process. Together they help identify the next question, rather than simply the highest score.
For Elena AILL, research into learning follows this connection: from what is retained to what the system understands and can use.
The work is ongoing. Its progress is recorded in tests that pass, tests that fail and the precision with which each finding informs the next step.
Sources and further reading
- Who is Elena? Why Elena AILL exists
The project's origin, founder, question about learning over time and stated limitations.
- F898 — F711 predicate realization
The comparison on 28 structured cases, the scope of the PASS and the unpromoted candidate.
- F896 — Meaning, knowledge and planner audit
The planner check and the insufficient end-to-end result, with separate denominators.
- Evidence-first methodology
How to read results, limitations and the boundaries of public evidence.
