Stored Is Not Reachable: Why Continual Fact Writing Breaks Under Later Updates

TL;DR for operators

A product team may repeatedly fine-tune a deployed model with new policies, prices, procedures, or customer facts, then approve each update because the newest item can be recalled immediately. That test shows the write changed current behavior. It does not show that earlier facts will remain usable after later updates.

After twenty sequential writes, facts trained with broad, varied study data retained 46% strict accuracy. Facts trained from bare statements retained only 1%.

The earlier information was often not simply erased. Even facts the model could no longer answer questions about retained 57% to 67% of the measurable increase in statement likelihood created during training. The write had left a statistical trace, but the question no longer provided a dependable route to it.

For operators, this distinction changes the acceptance test:

• Evaluate old facts after subsequent updates, not only immediately after training.
• Test uncued application and composition, not just repetition.
• Separate general capability regression from loss of access to specific facts.
• Use broad supervision when weight updates are necessary.
• Keep frequently changing, auditable, or jointly used knowledge externally retrievable.

Fine-tuning can successfully write a fact without making it durably reachable. Weight updates may improve performance, but they should not automatically become the sole record for knowledge that must remain reliable through continued change.

Immediate recall is a weak acceptance test

Consider a product team that periodically fine-tunes a deployed model on new policies or operating procedures. After each update, the team asks a few direct questions. The model gives the new answers correctly, so the update is approved.

That test establishes only that the latest training affected current behavior. It does not establish that the model can apply the fact in an unfamiliar situation, combine it with another new fact, retrieve it without a strong cue, or preserve it through future updates.

The paper isolates these differences using invented facts evaluated through five question types: recall, paraphrase, application, composition, and counterfactual reasoning. Questions were held out from training and retained only when the original model could not answer them. The evaluation also used a fact-in-prompt condition to verify that the questions were answerable when the required information was supplied.

This design prevents a direct-recall score from representing every form of knowledge use.

The gap appears even before continual updating begins. Training on a bare statement produced 97% recall and 95% paraphrase accuracy, but only 40% composition and 23% counterfactual accuracy. Broader study data produced lower raw recall—88%—but raised composition to 60% and counterfactual performance to 50%.

The bare-statement model had learned wording more successfully than use.

Broad supervision changes the kind of knowledge produced

One explanation for the study-data advantage would be that the broader material simply contains answers to the evaluation questions. The paper addresses this with a controlled factorization that varies prompt breadth while holding the model, fact set, adapter rank, learning rate, optimization steps, and seeds fixed.

Diverse recitation alone reduced the difference between lenient and strict scoring from 27.4 to 5.4 percentage points, without explicitly supplying derived conclusions.

The scoring distinction is informative. Lenient grading credits an answer that correctly states a premise from which the requested answer follows. Strict grading requires the model to provide the requested conclusion itself. A large gap therefore indicates that the model can repeat relevant information but does not reliably complete the inference demanded by the question.

The controlled result attributes much of that gap to data breadth rather than additional optimization or answer leakage. Repeating the same fact for more steps improved recall and paraphrase but did not substitute for varied linguistic forms and use conditions.

For an update pipeline, this changes the role of training-data generation. A canonical sentence may be sufficient when the sole requirement is direct retrieval under predictable wording. It is inadequate evidence for spontaneous use across paraphrases, decisions, or multi-step tasks.

Broad supervision does not solve continual learning, however. It produces a better initial write and improves its survival, but later access still deteriorates.

Later writes often break access before removing the trace

After twenty sequential fact writes, earlier facts trained with study data retained 46% strict accuracy. Bare-statement facts retained 1%. In the paper’s longer descriptive extension, study-written facts plateaued around 25%–28% after one hundred writes.

The principal retention comparison shows that initially usable knowledge is also more resistant to subsequent updates. The longer extension indicates that resistance is not the same as dependable accumulation; its design changed during the experimental process and its in-context ceiling missed the certification threshold, so it should be read as descriptive rather than a definitive long-horizon estimate.

The more revealing result comes from reconstructed checkpoints. The researchers measured how much of each write’s increase in the probability of its original statement remained after later updates. Even among behaviorally forgotten facts, the median drift-corrected retained lift was 57% for bare-statement facts and 67% for study facts after twenty writes.

This measurement is not a complete map of factual representation. It tracks the likelihood of the written sentence, not every semantic form the fact might take. Even with that boundary, it contradicts a simple interpretation in which failed question answering means the earlier weight change has disappeared.

Other probes point in the same direction:

Together, these findings support a functional account: the fact can leave a persistent trace while the model loses a dependable mapping from a question to that trace. Context improves performance because it supplies the needed information directly at inference time rather than requiring the model to reconstruct it through a degraded internal access pathway.

Incoming updates cause most of the measured interference

A team might respond by improving how each original fact is stored. The crossed experiments indicate that this targets only part of the problem.

The researchers separately varied two factors: whether an earlier fact had been stored using bare-statement or study training, and whether subsequent incoming facts used bare-statement or study training. Changing the incoming updates to study training improved final retention by 37.6 percentage points, with a 95% confidence interval from 28.9 to 46.3 points.

The earlier storage method had an estimated effect of −2.4 points, with a confidence interval from −9.2 to 4.1. In that experiment, it was not detectably different from zero.

This resolves a confound in ordinary sequential comparisons. If study-written facts survive better in a sequence where every later fact is also study-written, the advantage could come from stronger original storage, less damaging later updates, or both. The crossed design identifies the incoming update method as the main source of the measured difference.

The practical target therefore shifts from protecting each stored fact in isolation to controlling the entire update stream. Every new write should be evaluated for its effect on previously validated knowledge, not only for its success on the incoming item.

The paper also tested more direct controls. A local gradient-based predictor tracked the immediate effect of one update reasonably well, with a correlation of 0.795, but failed to predict long-horizon forgetting, where the correlation fell to −0.258. Bridge data, activation interventions, and protected gradient projection did not meet their prespecified success thresholds.

These are informative null results, not proof that interference cannot be controlled. They show that local compatibility with an earlier fact is insufficient to forecast a long sequence of optimizer steps and interacting updates.

Preserving general capability does not preserve earlier facts

Repeated writes can also degrade the model outside the newly trained facts. Across experimental conditions, general-capability loss was strongly associated with divergence from the original model’s output distribution. In the sixteen-condition factorial, the reported Spearman correlation was 0.946.

That makes distributional drift a useful diagnostic coordinate: large behavioral movement away from the base model often accompanies broader capability damage.

It is not a sufficient control objective. An explicit penalty could restore capability without systematically reducing endpoint divergence, and improvements in general benchmarks did not establish reliable access to earlier written facts. The paper therefore separates two deployment risks that are often combined:

1. Has the update damaged the model’s general behavior?
2. Can the model still retrieve and use previously written knowledge?

A frozen original-model teacher reduced capability damage compared with a moving teacher that accumulated prior writes. Recursive distillation through an already changed teacher amplified drift and produced the worst capability result among the compared conditions. A stable teacher is consequently a defensible control for general regression.

It does not solve factual reachability. A model can remain broadly capable while becoming unreliable on specific earlier writes.

Use weights as an optimization layer, not the sole knowledge record

The paper directly supports three controls for factual-update pipelines: broaden the supervision used for each write, retain a frozen reference model during repeated distillation, and validate old facts after every update sequence.

Cognaptus draws a narrower architectural inference. When information must be auditable, frequently revised, jointly reasoned over, or removed on demand, the canonical record should remain outside the model. Retrieval can then provide the required facts explicitly, while weight updates are reserved for behavior, style, stable domain regularities, or latency-sensitive caching where occasional loss can be independently detected.

This recommendation depends on the user and condition. A team adding a small set of stable product terms may accept weight writing when direct recall is the principal requirement and the update stream is tightly controlled. A compliance, support, finance, or operations team maintaining changing policies should not treat successful fine-tuning as proof that those policies will remain accessible after later updates.

A practical release framework therefore needs two gates:

Monitoring should also include newest-fact intrusion, looping, truncation, and degradation in the fact-in-prompt condition. In the experiments, some bare-statement sequences damaged generation even when the relevant fact was explicitly supplied, so apparent factual forgetting can coexist with broader output pathology.

The result is controlled, not universal

The evidence is comparatively strong inside its experimental setting: certified evaluation, paired comparisons, multiple seeds, controlled factorization, crossed interference designs, and selected replications at a larger model scale or with full-parameter fine-tuning.

Its production boundary remains substantial. Most experiments use Qwen3-4B and synthetic, discrete, stylistically homogeneous facts. Only the entailment-gap result was replicated at 8B scale. Several causal tests use small numbers of unique facts with repeated measurements. The joint-use and recovery experiments are also narrow, and the exact base-model revision was not recorded in the run rows.

The paper establishes neither that every natural enterprise fact will behave similarly nor that all weight-writing techniques will fail. It establishes that, under controlled continual updates, immediate recall substantially overstates what has been achieved. Broad training improves usability and retention, but no tested method provides reliable long-term reachability.

A successful write is only the first checkpoint

Continual factual learning requires at least three properties: the new fact must be usable, the model’s broader capabilities must remain intact, and earlier facts must stay reachable after future updates.

The experiments show that these properties can separate. A model may recite a new fact without applying it. It may stop answering questions about an earlier fact while retaining much of the write’s statistical trace. It may preserve general benchmarks while losing access to accumulated knowledge.

For operational teams, the acceptance criterion should follow the failure mode. Do not approve a factual update because the latest fact can be repeated. Test whether it can be used without cues, composed with other facts, and recovered after the next series of writes. Where those properties are mandatory, keep the source of truth externally retrievable.

Cognaptus: Automate the Present, Incubate the Future.