Learn AI /Model Quality and Evaluation /Memorisation
Memorisation
- Intermediate
- Model Quality and Evaluation
- Policy officers
- Technical readers
Also called: memorization, regurgitation.
Back to the milestone you came from
In five seconds
What it is
Memorisation happens when a model encodes information specific to individual training examples or small subsets of them.
In plain words
A model has memorised when it seems good mainly because it has seen very similar examples before, not because it learned a rule that works more widely.
An everyday example
Someone learns to sort reusable bottles by repeatedly seeing the same batch. Instead of noticing the bottle shapes and materials, they remember that the bottle with a torn green label goes in one crate and the bottle with a dent near the base goes in another. They are almost perfect when the familiar bottles return. With a new batch, the labels and dents are different, so the sorting quickly breaks down. The person learned details of the examples, not a rule that travels.
A person can usually explain what they memorised and may deliberately choose to do so. A model does not usually store examples as neat labelled photos it can point to, and its behaviour comes from many numerical parameters rather than conscious study habits. Also, some reuse of training patterns is normal and useful; the problem is harmful reliance on specific seen details that does not transfer to new cases.
Remember thisA strong score may reflect recalled examples or narrow cues rather than a pattern that transfers to genuinely new cases.
Understand it
Understand it
Why it matters
Memorisation matters because a system can appear reliable while actually being fragile. A polished dashboard or a high benchmark score may hide the fact that the model only works on cases that closely resemble its training data. For policy and operational use, that can lead to poor procurement choices, unfair treatment of some groups, weak risk controls and false confidence in deployment. Checking for memorisation helps you ask better questions: Was the test data truly separate? Are there duplicates or leaked clues? Does performance hold across places, languages, time periods and conditions that matter in the real task?
Learning a general rule versus memorising training examples
Show why two models can score well on familiar data but differ sharply on truly new data.
01
Same training set
Both models are trained on the same labelled satellite images of roads, including damaged and undamaged sections.
Top row spanning both columns: small thumbnails of labelled road images feeding into two separate model boxes.
02
General pattern learner
The left model picks up broad features that can appear in many places, such as broken road edges, irregular dark pits and repeated crack shapes.
Left column middle: model box highlights several distributed features across many thumbnails, with no single image emphasised.
03
Memorising learner
The right model leans on specific seen details, such as one neighbourhood pattern, one lighting condition or one repeated background object.
Right column middle: model box highlights a few exact thumbnails and a callout around a spurious cue like roof colour or shadow shape.
04
Looks fine on familiar checks
If evaluation images are too similar to training data, both models can appear accurate.
Third row: both columns output many green ticks for a test set labelled 'very similar to training'.
05
Separates on new data
When the test set comes from new cities, dates or image conditions, the general-pattern model holds up better while the memorising model drops sharply.
Bottom row: left column shows mostly green ticks on 'new images'; right column shows a visible mix of red crosses and a lower score badge.
A side-by-side comparison with three rows: training data, test on familiar-looking data, and test on genuinely new data. Read left to right. The left column shows a model that learns broad signs of road damage. The right column shows a model that relies on repeated image quirks from the training set.
Hovering or tapping a highlighted cue reveals whether it is a transferable feature or a training-specific shortcut.
Describe this without the diagram
This is a side-by-side comparison of two models trained on the same road images. The left model learns broad signs of damage that appear across many examples. The right model relies on specific details from the training images, such as a repeated background pattern or lighting cue. On a test set that looks very similar to the training set, both models seem to perform well. On a genuinely new test set with different places and conditions, the left model keeps performing relatively well, while the right model performs much worse. The diagram teaches that memorisation can hide behind good-looking results if the evaluation data are not truly independent and varied.
With reduced motion: A static two-column diagram shows the same comparison without animation: shared training data at the top, model behaviour in the middle, and performance on familiar versus new test data at the bottom.
A model is trained on examples such as images, text or numbers. During training, it adjusts numerical parameters to produce outputs such as a label, score or summary.
Memorisation occurs when those parameters encode information specific to individual training examples or small subsets. This may allow exact or near-exact reproduction, or make an output depend unusually strongly on whether a particular example appeared in training.
Memorisation can coexist with learning broader patterns. It can contribute to overfitting, but a drop on new data does not by itself prove memorisation. Such a drop may instead reflect distribution shift, noise or reliance on a spurious correlation.
Memorisation is also different from long-term memory in AI systems. Long-term memory is a designed mechanism for storing and retrieving information across separate sessions. Training-data memorisation is information encoded while the model is trained.
How “Sorting bottles by their dents” maps onto the real thing
| In the story | In the AI system |
|---|---|
| memorising particular practice photos and nearby visual quirks | storing training examples or their distinctive details rather than learning general features |
| high score on repeated practice images | strong performance on training data or leaked test data |
| mistakes on new images from other places and conditions | poor generalisation to unseen data |
| blue roof or curved shadow used as a shortcut | spurious cue that happens to correlate in training data |
A worked example
An information management officer at a regional hub checks an image classifier before using its results in a road-condition monitoring report.
Starting from: 1,200 labelled satellite images for training and 300 images for initial testing.
- Many images show the same road segments on nearby dates, creating near-duplicates with similar shadows, buildings and surface markings.
- A random split places related images in both sets. The model reaches 98% accuracy on the training set and 95% on the initial test set.
- The officer assembles a second test set from different areas, dates and viewing angles, removing duplicate and near-duplicate images.
- Accuracy falls to 68%. Errors cluster where shadows, surface colours and nearby structures differ from those in the training images.
Result: The initial score was inflated by repeated patterns and narrow visual cues. The drop on independently sampled images is consistent with memorisation and weak transfer, although further tests would be needed to establish the cause.
In a United Nations workplace
A partnerships officer at a shared services centre is finalising a handover note before close of business. She reviews a model proposed to route incoming implementing-partner queries into categories such as grant amendments, reporting and due diligence.
The historical inbox contains repeated templates, familiar partner names and near-duplicate follow-ups. If these appear across both training and test sets, the model may score well by recalling narrow cues rather than learning the substance of each query. It may then misroute messages from new partners or a new reporting cycle.
Before you try thisDo not place correspondence containing personal data, bank details, due diligence records or procurement-sensitive material in any tool not approved for it. Use an approved environment, restrict access, remove unnecessary identifiers and duplicates, and test on later, independently sampled messages. Human review remains necessary because both disclosure and misrouting can cause harm.
Where you may hear it
- During first reading of a model evaluation report
- At a vendor demo in a procurement meeting
- In an ICT governance board review of a proposed pilot
- During clearance of a monitoring report that relies on model outputs
Avoid confusion
Avoid confusion
Often confused with
Overfitting
A model fits its training examples so closely that its performance drops on new cases.
Generalisation
The ability of a model to give useful results on new cases, not just on the examples used to build or tune it.
Long-term memory (in AI systems)
Stored information an AI system can retrieve across separate sessions to personalise, continue tasks, or reuse past facts and preferences.
Common misconceptions
Common misconception: If training accuracy is very high, the model must have learned the task well.
Correction: High training accuracy only shows the model fits the data it has already seen. It may still rely on repeated examples, accidental clues or noise. What matters is performance on well-separated, representative new data.
Common misconception: Memorisation and overfitting are exactly the same thing.
Correction: They are closely related but not identical. Overfitting is the broader problem of fitting too closely to training data and losing performance on new cases. Memorisation is a specific pattern within that broader problem, where the model depends heavily on seen examples or highly specific details.
Common misconception: Memorisation only matters for text models that repeat training sentences.
Correction: It can happen in images, tabular data, audio and other formats too. A model may memorise repeated backgrounds, camera artefacts, site names, timestamps or other shortcuts, even if it never reproduces an exact text passage.
Common misconception: A good benchmark score proves memorisation is not a problem.
Correction: Not necessarily. A benchmark can still be too narrow, too similar to the training data, or affected by contamination and leakage. The score is useful only if the evaluation setup matches the real use case and is properly separated from training.
Quick check
Which result is the clearest warning sign of memorisation?
High scores on training data and near-duplicate test data can happen even when the model has mainly learned to repeat seen patterns. The warning sign becomes stronger when the test set is not truly independent. Good evaluation uses separate, representative data and checks whether performance holds under meaningful changes.
Go deeper
Go deeper
The simple version
A model has memorised when it seems good mainly because it has seen very similar examples before, not because it learned a rule that works more widely.
The technical version
In technical terms, memorisation refers to a model encoding training-set-specific information strongly enough that its outputs depend on recalled examples or narrow cues rather than robust, transferable structure. This can appear as exact or near-exact reproduction, unusually high confidence on repeated patterns, sensitivity to duplicate or near-duplicate samples, or sharp performance drops under distribution shift. It is often associated with high-capacity models, limited or noisy data, and weak train-test separation. Memorisation is related to but distinct from generalisation: some retention of recurring patterns is necessary for learning, while harmful memorisation shows up when behaviour does not extend to independently sampled data from the target task. It is also linked to overfitting and can be worsened by data leakage, contamination of benchmarks, or spurious correlations. Detection often relies on holdout evaluation, deduplication, cross-validation where appropriate, out-of-distribution testing, privacy-oriented probing, and inspection of whether errors increase when superficial cues change.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This official NIST framework gives a risk-management lens for evaluating AI systems, and it suits policy and governance readers who need to connect technical issues like memorisation to assurance practice.
Intermediatelink checked 2026-07-31
EncyclopaediaWikipedia search: Memorisation (opens in a new tab)Wikipedia
This Wikipedia search results page helps readers find background material on the term from multiple angles, and it suits readers who want a broad starting point rather than one fixed article.
search results, not one pagelink checked 2026-07-31
CourseClassification: Accuracy, recall, precision (opens in a new tab)Google
This course resource explains core classification measures such as accuracy, recall and precision, and it suits readers who need context for spotting when good-looking scores may hide memorisation.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Memorisation (opens in a new tab)YouTube search
This YouTube search results page offers a range of video explanations on the term, and it suits readers who prefer visual or spoken introductions and are willing to compare sources.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Memorisation (opens in a new tab)Reddit search
This Reddit search results page surfaces community discussions and examples of the term in practice, and it suits readers who want informal explanations and should be read critically because it is a search page of mixed-quality posts.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A strong score may reflect recalled examples or narrow cues rather than a pattern that transfers to genuinely new cases.