Research topic · Evidence-first

Continuous learning in artificial intelligence: experiments and limits by Elena AILL

Continuous learning in artificial intelligence studies how a system can acquire or update knowledge over time without losing what it still needs. It is different from training performed only once: after each update both the new behavior and the preservation of the previous one must be verified. Learning one new item is not enough, and a single PASS does not demonstrate general learning. Elena AILL is presented here as a documented experimental case, with limited results, candidates not promoted and failures preserved alongside successes.

What is continuous learning in artificial intelligence

In general terms, a continuous learning system receives new information, examples or tasks at successive times and updates its behavior without necessarily starting from scratch. The evaluation does not only concern what is acquired in the last step: it must also check the previous capabilities, because a useful update on the new one can interfere with what the system already knew how to do.

“Continuous” does not mean that the system learns all the time, without control or without limits. Implementations can use online procedures, local updates, offline consolidations, isolated candidates, or promotion cycles. To correctly describe an experiment it is necessary to indicate which state changed, which data were used, which gates were applied and whether the result remained a candidate or was promoted.

In Elena's case, the phases adopt different procedures. Some add facts and verify the recall; others measure feedback influence, generalization, anti-forgetting or selection of a candidate. This page does not turn these tests into a universal protocol and does not state that Elena learns independently in an open environment.

Learning is not just about remembering

Persistent memory mainly concerns conservation, availability after a save or reload and recall ability. Continuous learning requires something more: acquisition or updating, testing on the new, retention of the previous and evaluation of the interferences introduced by the change.

A fact recalled after a reboot may demonstrate persistence within the test perimeter, but it does not alone demonstrate that the system has generalized, improved a skill, or retained all prior knowledge. Likewise, an update that improves a new set isn't enough if it degrades previous sets.

Memory taxonomy, cold reload and persistence tests belong to the dedicated hub persistent memory and recall after reload. Here those results are only recalled when they serve to evaluate retention after an update.

Continuous learning, lifelong learning and incremental learning

Term	Use on this page	What it doesn't imply
Continuous learning / continuous learning	Sequential acquisition of the new with verification of previous knowledge.	Continuous, gateless, or universal learning.
Incremental learning	Progressive updating through examples, experiences or subsequent tasks.	That every update is promoted in runtime.
Lifelong learning	Research direction towards accumulation and reuse of knowledge along extended sequences.	Infinite learning or definitive absence of forgetting.
Live learning	Update or acceptance checked during a candidate operational path.	Global training open during every conversation.
Local learning	Circumscribed modification driven by local signals and verified in the relevant test.	Update of the entire system or causality on the final text.

The historical names of the phases Elena — how Continuous Learning, Live Learning Runtime and Lifelong Memory Expansion — remain unchanged. The name of a phase describes its experimental objective; it does not replace the status, verdict and measures recorded.

The problem of catastrophic forgetting

Catastrophic forgetting occurs when the acquisition of new information or tasks markedly degrades previous performance. The difficulty can be read as a balance between plasticity, that is, the ability to change and acquire the new, e stability, that is, the ability to preserve knowledge and behaviors that are still useful. Too much plasticity can produce interference; too much stability can prevent adaptation.

The Elena phases use different controls: recall of the old and new, behavior probe, regression gate, anti-forgetting score, reload or paraphrase test. These values are valid only for the declared sets. For example, F301 records 10/10 previous recalls and four new facts in its test; F608, F609 and F610 they report candidate measures at increasing scales, but also state that the candidate was not promoted. None of these results demonstrate that Elena has “solved” catastrophic forgetting.

How Elena tests new learning

The procedures change from phase to phase, but the documented evidence allows us to recognize a recurring conceptual sequence. Not all steps appear in every experiment.

Baseline or initial state: previous behavior, recall, checkpoints, or counts are recorded.
New information or update: a fact, an experience, a local signal or a candidate set enters the expected path.
Verification of acquisition: Direct probes check whether the new item is available or influences the measured behavior.
Previous check: old-set recall, smoke tests or regressions check that previously recorded capabilities have not been lost.
Generalization: when expected, paraphrase, non-identical properties or cases verify a transfer beyond the literal recall.
Persistence: some phases save a candidate and verify reload; others explicitly state that checkpoints or frozen runtimes have not been touched.
Gate and promotion: measurements are compared with phase criteria; a candidate can be retained, rejected or not promoted.
Final classification: PASS, FAIL, PARTIAL, MIXED and “Validated without terminal verdict” remain distinct categories.

Documented evidence on continuous learning

The following selection covers acquisition, live/local learning, retention, generalization, scale and non-promoted candidates. It is not a list of victories only and does not replace the individual phase pages.

Phase	What head	Outcome	Main measurement	Limit
F61-B	Self-learning speech without declared supervised labels	Documented; no terminal verdict	Dataset 1,050,147 bytes; active queue; pseudo-word ratio 0	It is neither a terminal PASS nor a runtime test.
F96	New facts and previous behavior	PASS	New facts 10/10; probe 15/15; old behavior 5/5	Controlled set, not open learning.
F268	Feedback and limited repair	PASS	3/3; repair used 3 times; training not started	Local loop, not general continuous learning.
F300	Controlled live learning and reload	PASS	4 new facts after reload; useful experiences accepted and noise rejected	Unchanged base and isolated candidate path.
F301	Retention after live learning	PASS	Old recall 10/10; 4 new facts; anti-forgetting 1	Score relating to the phase set.
F514	Autonomous Learning Loop candidate	PASS	State score 0.4366; 5 intents; training started	The name does not imply general autonomy.
F515	Candidate loop on test data	PASS	State score 0.962; accepted ratio 1; 7 intent	It does not test open-world robustness.
F607	Teach surface and recall after reload	Validated; no terminal verdict	Reload recall 1; teach-mode pass = no; not promoted	A negative subgate prevents it from being called a full PASS.
F608	Lifelong candidate 1K	Validated; no terminal verdict	Direct/paraphrase/anti-forgetting 1; at least 900 concepts saved	Candidate not promoted.
F609	Lifelong candidate 10K	Validated; no terminal verdict	At least 1,000 cases; recall and anti-forgetting 1	Candidate not promoted.
F610	Lifelong candidate 100K	Validated; no terminal verdict	At least 1,000 cases; recall and anti-forgetting 1	It does not demonstrate unlimited lifelong learning.
F684	Lifelong expansion and retention	PASS	100,206→150,236 concepts; direct recall and anti-forgetting 1	Counting concepts, not parameters or global intelligence.
F698-C	Activating candidate components	PARTIAL	Stack used 4 times; active lifelong/feedback/association gates	It is not complete ordinary integration.
F879	Causal learning and paraphrasing	PASS	Direct and property recall 1; paraphrase ≥90%; anti-forgetting 1	Generalization limited to the recorded test.

When learning fails

Negative evidence delimits what positive results do not allow us to affirm. One component can pass its check while the overall phase fails.

Phase	Test	Result recorded	Because it limits the claim
F65-H	Real update and generalization across 1,000 prompts	FAIL; gate pass = no	Local recall is not enough when generalization does not pass the gate.
F105	Blind failure learning on 99 cases	FAIL · PASS; phase failed	A retest or PASS component does not convert the terminal result to success.
F700	True blind freeze on 500 cases	FAIL · PASS; phase failed	100% used stack and self-eval do not avoid end-to-end failure.
F701	Failure Driven Corpus	FAIL · PASS; phase failed	The local target passes, but the stack is not used and the candidate is not promoted.
F705	5, 20, 50 and 100 turn continuity	MIXED; FAIL_DO_NOT_PROMOTE	References are not retained and non-promotion remains part of the result.

These failures do not prove that “the whole system is broken”. Instead, they demonstrate that claims must remain tied to the level actually exceeded and that integration, generalization and continuity require separate proof.

Continuous learning and persistent memory

Persistent memory is necessary in many continuous learning protocols because it allows state or knowledge to be preserved between steps and reloads. However, it is not enough: a system can retain data without acquiring a new capability, and it can acquire something new by degrading what it had learned before.

For this reason, the most informative tests combine new recall, old-set retention, regressions, possible generalization and the candidate's status. Full explanations on memory types, cold reload, word memory and 10M benchmarks remain in the hub on persistent memory in artificial intelligence.

Limits and risks of continuous learning

Forgetting and interference: the new can degrade previous knowledge or behavior.
Contamination: Noisy, malicious, or out-of-context data can introduce unwanted associations.
Incorrect information: persisting an update does not make it true; Provenance, correction and rollback are needed.
Drift: a sequence of small changes can move the system away from validated behavior.
Incomplete rating: a restricted set may not detect regressions, shortcuts, or out-of-distribution fragilities.
Promotion risk: candidate and promoted runtime must remain distinct; non-promotion and frozen state were significant results.
Privacy: when new data persists, collection, provenance, isolation and deletion must be considered. The steps mentioned do not demonstrate complete public coverage of these controls.

Elena documents some isolation gates, checkpoints, candidate promotion and regression. This page does not grant universal protections where the archive does not provide specific verification.

How to read Elena's experiments

Status does not mean verdict: “Validated” describes the classification of the record, it does not automatically equate to PASS.
Candidate does not mean promoted runtime: a candidate can pass limited tests and remain not promoted.
Component PASS does not mean end-to-end PASS: F105, F700 and F701 retain PASS components within failed phases.
A perfect score belongs to the defined set: it does not demonstrate a general absence of forgetting or arbitrary generalization.
A failure is local to the test: it limits the tested claim, it does not prove that every other part of the system fails.
Training, checkpoint and reload should be read separately: some phases change candidate state, others declare training not started or frozen checkpoint intact.