AI memory degradation is real.

We tend to judge an AI by the answer it gives right now. Can it solve the problem? Can it explain the idea? Can it write the code? An assistant that remembers us raises a different question: can it still make sense of what came before?

I started with a more provocative version of that question: can AI develop Alzheimer’s? Taken literally, the answer is no. Alzheimer’s is a biological disease, and a language model has no biological brain. But that does not make the concern behind the question imaginary. [4]

Researchers have already measured AI systems losing previously learned abilities, struggling to use information still available to them, and producing unreliable records of earlier interactions. These are observed failures, not predictions about a distant future. [1] [2] [3]

The unsettling part is that the underlying model does not always have to change. Its intelligence can remain intact while the information it relies on becomes harder to reach—or less trustworthy.

Part of the confusion comes from the word “memory.” It can mean what a model learned during training, the material placed in its current context, or the records an application saves and retrieves later. These are different systems. They can fail for different reasons.

A wrong answer alone does not prove deterioration. The useful evidence comes from comparisons: performance before and after new training, recall with and without a long surrounding history, or behavior before and after stored experience becomes corrupted. That is where the research gets interesting.

When learning something new costs something old

Catastrophic forgetting is an established machine-learning phenomenon. Kirkpatrick and colleagues’ 2017 PNAS paper demonstrated that sequential learning can disrupt earlier tasks and that selectively constraining important weights can improve retention. Their experiments included handwritten-digit classification and Atari games. This is direct evidence about learning and retention, not a metaphor derived from human disease. [5]

Luo and colleagues tested continual instruction tuning in language models. For BLOOMZ‑7.1B, all three general-capability aggregates below were lower at the end of the five-task training sequence. These are the paper’s before-and-after scores, not its separate average forgetting metric. [1, Table 4]

These results demonstrate loss after additional learning in the tested models and training setup. They do not establish equivalent losses for every architecture, training recipe, or current commercial model. The important causal distinction is that the weights were updated; this is not evidence that ordinary inference wears them out.

The information is there. The answer still slips away.

In Lost in the Middle, published in Transactions of the Association for Computational Linguistics in 2024, Liu and colleagues varied where answer-bearing information appeared in a model’s input. Across multi-document question answering and key–value retrieval, performance was often strongest near the beginning or end and weaker in the middle. The relevant information had not been deleted; its position affected whether the model used it successfully. [6]

This is evidence of an access-and-use limitation, not physical memory damage. It also makes a larger advertised context window an insufficient guarantee of reliable recall: accepting an input and extracting the right information from it are separate capabilities.

LongMemEval, published at ICLR 2025, makes the comparison especially concrete. In its no-Chain-of-Note condition, the same questions were answered using only their evidence sessions (“Oracle”) or the full approximately 115,000-token history (“S”). All five model configurations below performed worse on the full history. [2, Figure 3b]

The contrast supports degraded use of available conversational information without retraining. It does not show that a specific stored record was destroyed or that the model aged.

A saved memory is not necessarily a reliable one

The HaluMem preprint evaluates external-memory extraction, updating, and answering. Its Long condition adds irrelevant dialogue to Medium histories. Answer accuracy fell for all four systems below, but by markedly different amounts. [3, Table 3]

Supermemory was nearly stable, and its extraction recall improved. The authors report upstream errors propagating into answers. Limitations include generated histories, automated judging, and authors affiliated with a memory-system vendor. Independent replication matters.

The 2026 MemGuard preprint identifies a specific mechanism: stable facts, episodic events, and behavioral rules can be mixed and used as if they were equivalent evidence. A context-specific event can then become an overgeneralized claim. Its type-aware approach reported improvements in memory construction and retrieval. [7]

However, its HaluMem comparison changes both the underlying model and the judge relative to the imported baselines, so it is not a clean, single-variable architectural comparison. Its own generation results still include hallucinations and omissions. The justified conclusion is that memory contamination is an observed failure mode—not that one architecture has eliminated it.

These findings shift the issue from “the assistant forgot” to a more precise question about the record: was the information captured, preserved, updated, retrieved, and interpreted correctly? Those are distinct stages. A correct original conversation does not automatically guarantee a correct derived memory.

More conversation can make things worse

Laban and colleagues compared fully specified, single-turn instructions with the same requirements disclosed across several turns. Across more than 200,000 simulated conversations, 15 models, and six generation tasks, they reported an average 39% relative performance drop in the multi-turn setting. Concatenating the disclosed requirements into one prompt largely preserved performance, helping distinguish conversational effects from simple information loss during rephrasing. [8]

Their analysis linked the result primarily to increased unreliability, including premature assumptions and dependence on earlier flawed attempts. This is measured conversational degradation, not a claim that 39% of memories were erased. The tasks include coding and other generation tasks, so the number is not a general memory-recall score.

The result strengthens a narrower point: an unchanged model’s useful behavior can worsen because of the trajectory of its interaction. It does not demonstrate weight decay. The study originated in 2025 and is listed by Microsoft Research as an ICLR 2026 publication. Its tested model versions, simulation, and tasks bound the finding.

When AI starts learning from its own echoes

Shumailov and colleagues’ 2024 Nature paper demonstrated model collapse in recursive generative-training regimes. Successive models were trained on outputs from earlier models; distributional errors accumulated and information about the original distribution was lost. Their language-model experiments used OPT‑125M and WikiText‑2, with worsening behavior across generations. [9]

This is evidence of degradation through a learning feedback loop. It is not evidence that every AI-generated training example is harmful, or that a frozen model summarizing a chat undergoes the same process. Training changes learned representations; summarizing changes the information supplied to a later inference. Treating the two as identical would overstate the paper.

Gerstgrasser and colleagues found that retaining original data while accumulating later synthetic data avoided collapse in their studied settings, unlike repeatedly replacing the dataset. Their experiments included language models, diffusion models, and variational autoencoders, alongside a theoretical analysis of linear models. [10]

In the language-model experiments, the original corpus was TinyStories—itself synthetically generated—so “original” should not be casually equated with “human-authored.” The relevant distinction is retaining the initial reference distribution rather than repeatedly replacing it with descendants’ output.

These results do not refute the existence of collapse. They refute an unconditional version of the claim. A demonstrated failure under one feedback regime and stability under another indicate dependence on data management, not inevitable biological-style aging.

The past can also be poisoned

MemoryGraft examines a different cause: adversarial rather than accidental contamination. In experiments with MetaGPT’s DataInterpreter and GPT‑4o, the authors introduced poisoned experiences into persistent retrieval memory. Later tasks could retrieve and imitate those experiences, carrying the influence beyond the original interaction. [11]

This preprint supports the possibility of persistent behavioral corruption through external memory. It does not show spontaneous degeneration, and its agent implementation does not represent every deployed assistant. Nevertheless, it establishes why intact reasoning machinery is not enough: a system can act on a corrupted account of what previously worked.

From a reliability perspective, “memory” is therefore not automatically a trusted source just because the assistant wrote it. A retained mistake can be encountered repeatedly, and a record of an instruction is not legitimate authority to follow that instruction.

Intelligence is not the same as a trustworthy past

Taken together, these studies establish something more concrete than the Alzheimer’s analogy. Artificial systems can lose learned performance, become worse at using available information, and carry unreliable memories into later behavior. The evidence supports those failures under documented conditions, not a universal decay rate. [5] [6] [11]

I use Digital Cognitive Degeneration as a way to connect these concerns. It is my umbrella term in this article, not an established diagnosis or a claim to have discovered the phenomena. Catastrophic forgetting, memory contamination, and model collapse already have their own research histories.

Bringing them together does not mean treating them as one disease. Nor does it prove that every AI inevitably gets worse with age. A failed retrieval, a corrupted record, and a capability lost during training need different explanations. Better memory systems can also make a difference, as the counterexamples show.

What matters to someone relying on an assistant is whether it can give a dependable account of what happened before. A fluent answer is not enough. Neither is a large archive.

An AI’s ability to reason and its ability to remember faithfully are not the same thing. The research gives us good reason to stop treating them as if they were.