Forgetting in machines: when deleting data isn't enough

A datum erased from a table disappears. A datum erased from a trained model does not. It stays there, dissolved in the weights, diffuse but recoverable, altering every future response of the system without any query revealing it. This article begins where the previous one on forgetting in classic storage ends: once the datum leaves storage, it has already left traces in indexes, embeddings, caches and —hardest of all— in the model’s own weights. And against this technical landscape, human forgetting —biological, adaptive, functional— appears as the mirror that reveals what machines still don’t know how to do.

Three kinds of forgetting that are not the same

When people talk about “forgetting” in AI systems, three technically distinct phenomena get mixed up. Before proposing anything, they need to be separated, because confusing them hides the real problem.

Catastrophic forgetting. The oldest and best known. A model trained on task A that is then trained on task B tends to unintentionally lose what it knew about A. There is no intent to erase; it just happens, as a side effect of learning. A January 2026 paper on continual unlearning in LLMs confirms it: previously seen examples are gradually —or catastrophically— forgotten throughout training, and this is especially true for language models.

Machine unlearning. The exact opposite: forgetting induced on demand. When a regulation demands that a model “erase” the contribution of a specific datum —typically through a request under Article 17 of the GDPR— engineering needs techniques to remove that contribution without destroying everything else. It is a young and still fragile field.

Implicit retention in the weights. The least discussed and probably the most uncomfortable. A model trained on a datum retains information about that datum in its parameters even if the original dataset is deleted. Storage forgets; the model does not. A piece of personal data that entered a training pipeline can keep influencing the model’s responses throughout its useful life, without any visible record showing it.

The confusion between these three is not semantic. It is what allows many organisations to state “we have erased your data” when what they have done is erase the primary record and leave the already integrated contributions intact.

Forgetting as propagation, not as a single act

Here it is worth pausing on something the article on classic storage pointed to but did not develop: when an erasure request arrives in a modern AI system, deleting the primary record is the trivial step. What is hard is everything else.

An academic arXiv paper from 2026 on faults and pitfalls in the implementation of the right to be forgotten documents the pattern: modern systems are distributed architectures where personal data leaves derived traces across multiple layers. In a system with AI components, those layers are especially treacherous:

• Vector indexes with embeddings computed over the datum
• Semantic caches of RAG that retain fragments of the context
• Trace logs with historical prompts and responses
• Derived datasets for fine-tuning, validation or evaluation
• Trained models with weights updated by the datum
• Incremental backups that hold previous versions

An erasure at the primary layer can coexist for months with functional copies at subordinate layers. Forgetting, in a distributed AI system, is not a punctual act — it is a propagation that has to be orchestrated. And of all the layers, the last one is the hardest to handle: the model’s weights have no clean deletion.

The engineering of machine unlearning: the essentials

The state of the art has developed quickly, but the honest picture is that nobody yet has a solution that is robust and verifiable at the same time. It is worth seeing why, staying at the essentials.

The exact technique: verifiable but prohibitive

The only technique that mathematically guarantees demonstrable forgetting is retraining the model from scratch without the datum to forget. For a contemporary LLM this means millions of dollars, weeks of compute and a considerable energy footprint per request. It’s a solution that doesn’t scale.

A viable variant is SISA (Sharded, Isolated, Sliced, Aggregated): split the dataset into isolated fragments, train independent submodels per fragment, combine their outputs. When a request arrives, only the affected fragment has to be retrained. It’s an architectural solution: it bakes forgetting into the system design, not as a subsequent add-on. Its limit: it demands committing to the architecture before training, not after.

The approximate techniques: the fragile family

The bulk of recent research has focused on techniques that don’t retrain, but modify the existing weights to reduce the influence of the datum to forget. The most representative is Gradient Ascent, which increases the loss over the set to forget until the model stops predicting correctly on it. Other variants use isolated LoRA adapters, task vector negation or contrastive decoding at inference.

All are elegant. All have the same problem.

The problem nobody is solving well: reversibility

A May 2026 paper titled Unlearning Isn’t Deletion documents what research is starting to call fragility of forgetting: most current methods produce an apparent forgetting that can be reverted with relative ease. An “unlearned” model can recover the suppressed knowledge through light fine-tuning on seemingly innocuous data —so-called benign relearning—, through low-precision quantisation, or through adversarial prompting.

Translated: much of what today is called unlearning is actually a superficial suppression that hides the memory without eliminating it. The painful analogy is that the model learns not to say what it knows, not to not know it. That difference is exactly what separates real compliance from performative compliance.

And there is an additional dimension the public conversation barely touches: continuous maintenance degrades the model. In an environment of accumulated requests —continual unlearning—, each new forgetting operation interferes with the previous ones, producing drift in the parameters until the model can functionally collapse. A system committed to forgetting on demand is a system that degrades by design with each request it honours.

What this changes in data engineering applied to AI

Three concrete decisions emerge from this landscape.

Minimisation at training, not just at storage. If a piece of personal data should not be recallable later, it probably should not have entered training in the first place. Every integrated piece of personal data creates a future obligation that is hard to honour. And “hard to honour” here means expensive technique, impossible verification and progressive model degradation.

Modular architectures for localised forgetting. If total model forgetting is prohibitive and approximate forgetting is fragile, the viable exit is to design the architecture so that forgetting is localised. SISA is the best-known proposal, but the logic applies in general: LoRA adapters per data batch with independent discard, RAG with external memory that keeps personal data outside the model and allows deletion by index, and deliberate separation between base models trained without personal data and adaptive layers that do incorporate it.

Explicit acceptance of the limits. There are erasure requests that engineering today cannot fulfil with guarantees. A large model trained a year ago with data that wasn’t isolated by batches cannot cleanly forget a specific datum. Acknowledging this is a call to future architecture: models from here on must be designed assuming that on-demand forgetting is a requirement, not an optional feature added later.

What this approach does not solve

None of the current techniques solves the underlying problem: that generative models are systems whose intelligence emerges precisely from the massive integration of data, and that integration is intrinsically resistant to surgical removal. It is like asking a river to give back the drops a tributary contributed a year ago.

Nor does it solve the asymmetry of auditability: verifying that a model has forgotten demands access to the weights, compute and expertise that only the provider and a few researchers possess. Trust in forgetting depends, in practice, on the good faith of the operator.

The other side: human forgetting as a functional mirror

So far, engineering. But it is in the comparison with biological forgetting where this technical problem becomes something bigger — and where the magnitude of what machines still don’t know how to do is revealed.

Biological forgetting is not a failure

For decades, neuroscience treated forgetting as a deficit. Research over the last fifteen years has demolished that reading. Studies published in Nature and Nature Communications have shown that the brain intentionally suppresses specific memories when they compete with current retrieval goals, and that this process depends critically on the prefrontal cortex. The 2024 review in Annual Review of Psychology sums it up without ambiguity: memory exists to serve an adaptive purpose, and that purpose demands forgetting as much as remembering.

Biological forgetting has three documented functions: emotional regulation (it limits access to negative memories), knowledge acquisition (it enables generalisation by abstracting and automating), and contextual tuning (it orients processing towards the relevant present).

The key formulation, repeated in the literature with variants: a brain that retained everything with equal weight would not think better. It would not think at all. Forgetting does not compete with cognition; it enables it.

The asymmetry this mirror reveals

Placed in parallel with what machines do today, the contrast is uncomfortable:

The brain forgets by function; the model forgets by accident. Catastrophic forgetting does not respond to any criterion of relevance. What is forgotten is what gets stepped on when learning something new, regardless of whether it was important or trivial, useful or harmful.

The brain selects what to forget; the model does not distinguish. The prefrontal cortex decides, based on current goals, which memories to suppress. A current generative model has no equivalent: when forced to forget, it applies the technique homogeneously, without functional criterion.

The brain forgets in order to generalise; the model relies on retention to reason. Biological abstraction is the counterpart of forgetting specific details. A model that retains the detail without forgetting isn’t generalising — it’s memorising. And that memorisation is what makes it so hard to erase later.

A recent survey on forgetting in machine learning points to the most promising direction: integrating adaptive forgetting mechanisms inspired by the brain not as a compliance add-on, but as a design principle to improve the system’s own intelligence. Not forgetting to comply with the GDPR; forgetting because forgetting well is part of thinking well.

The responsibility of the builder

The useful question is not “how do we make machines forget like humans?” — it’s anthropomorphic and probably wrong. The question is: how do we design systems that integrate forgetting mechanisms with functional criterion, not just with regulatory pressure?

That means thinking about forgetting in three simultaneous dimensions: selectivity (forgetting the specific without erasing the structural), functionality (forgetting to improve the system, not to fulfil external obligations), and verifiability (demonstrating that the forgetting has occurred well). All three must be present. None alone is enough.

Open questions

• If a trained model retains information about personal data in its weights even when the data is erased, at what point can it honestly be claimed to have forgotten?
• When today’s unlearning techniques produce reversible forgetting through light fine-tuning, are we complying with the right to be forgotten or staging it?
• Can an AI system aspire to sustained intelligence without developing functional forgetting mechanisms, or is accumulative retention the asymptote that will stop it?
• If human forgetting is a function and machine forgetting is an accident, what are we really asking engineering when we demand that machines “forget like us”?

The questions have no closed answer. But one final idea is worth keeping: well-designed forgetting is not the absence of memory — it is a more mature form of memory. And that more mature form, today, is a territory where biological brains have millions of years of advantage over artificial ones. Acknowledging that asymmetry is not humility; it is the precondition for building systems that may one day close the gap.

Unlimited digital memory is not an achievement. It is a shortage disguised as a capability. And that distinction belongs to whoever sustains it with their code.

References

• arXiv — Wuerkaixi et al. FIT: Defying Catastrophic Forgetting in Continual LLM Unlearning (January 2026). arxiv.org/html/2601.21682v1
• arXiv — Rezaei et al. Unlearning Isn’t Deletion: Investigating Reversibility of Machine Unlearning in LLMs (May 2026). arxiv.org/html/2505.16831v3
• EmergentMind — Machine Unlearning for LLMs (February 2026). emergentmind.com
• Bourtoule et al. — Machine Unlearning (SISA architecture, 2021). The original architectural proposal.
• arXiv — Faults and Pitfalls in Implementing the Right to be Forgotten (2026). arxiv.org/pdf/2605.27171
• Nature — The forgotten part of memory (2019). Forgetting as an adaptive function. nature.com
• Nature Communications — A retrieval-specific mechanism of adaptive forgetting in the mammalian brain (2018). nature.com
• Annual Review of Psychology — Anderson & Hulbert. Active Forgetting: Adaptation of Memory by Prefrontal Control (2024). annualreviews.org
• arXiv — “Forgetting” in Machine Learning and Beyond: A Survey (2024). arxiv.org/html/2405.20620v1
• Previous article in this series: Infrastructures that forget.