Catastrophic Forgetting in LLMs: What happens when you retrain?

Continuous learning in LLMs sounds like an obvious function: import new data, briefly retrain, update knowledge. This is exactly where the operational risk begins. Fine-tuning can improve performance on the new training data while degrading skills that were reliable before the update. This Catastrophic Forgetting does not arise from the Transformer architecture alone. It is a risk of gradient-based, sequential optimization when new training signals compete with already learned parameters.

The knowledge cutoff is to be separated from this. It initially arises because a basic model is trained on a completed data set and then delivered as a fixed version. Further training is technically possible. What is difficult is to incorporate new knowledge in a targeted, affordable and verifiable manner without uncontrollably damaging existing knowledge or skills. The preprint “Language Models Need Sleep” examines this conflict in 2026 with Qwen models. His suggestion is currently research, not a ready-made update function for productive LLMs.

Continual Learning in LLMs: Why Forget Gradient Optimization

A neural network learns by gradually shifting weights toward minimal loss on the current training batch. This works well as long as the data stream remains stable. But as soon as a new task or a new data set appears, the gradient moves in a different direction. Weights that were important for the old task are also carried along. The newer signal can displace the older one.

This problem has been described in classical neural networks since the late 1980s. It is particularly relevant for LLMs because general language ability, factual knowledge and task-specific behavior are distributed across many parameters. A small specialized data set can strongly pull the optimization towards the new data. Whether and how much the model forgets depends on the training data, the learning rate, the updated parameters and the training method.

This results in a clear obligation for operators to test: fine-tuning must also be tested against the capabilities that should be retained.

Catastrophic Forgetting: What EWC can protect

An influential classical approach comes from James Kirkpatrick and colleagues. Her paper “Overcoming catastrophic forgetting in neural networks” was published in 2017 Elastic Weight Consolidation (EWC) a. The method estimates which weights are important for a previous task and penalizes large changes to these parameters when learning a new task.

EWC uses the Fisher information matrix to estimate, for each parameter, how sensitive the loss of the old task is to its change. Weights with high sensitivity are slowed down more during the new training. In the Atari experiment, Kirkpatrick et al. show that an agent could learn multiple games sequentially without completely forgetting the first one and without using separate network instances.

The basic idea remains relevant, but does not automatically scale to billions of parameters. Importance estimates require additional memory and computational effort. In addition, any regularization protects the old at the expense of some of the adaptability. This is the stability-plasticity dilemma in practical form.

An LLM-specific work by Song and colleagues from 2025 therefore uses hierarchical regularization at the layer and parameter level. The authors test GPT-J and Llama 3 on scientific, medical and physics tasks. They report less forgetting and significantly lower memory requirements than previous comparison methods. However, the finding comes from a “work in progress” preprint and is not general proof for every fine-tuning setup.

Solutions for new knowledge

A 2025 ACM Computing Surveys overview classifies Continual Learning in LLMs along Continual Pretraining, Instruction Tuning and Alignment. A broader distinction is more useful for operators: RAG, fine-tuning, model editing and continuous pretraining solve different problems and are therefore not interchangeable.

RAG (Retrieval Augmented Generation) is the most pragmatic way when it comes to knowledge updating. Instead of changing the model itself, an external document storage is connected from which the model draws relevant passages at inference time. The model doesn’t learn anything new; it is provided with context. RAG is well suited for domain-specific factual knowledge and is by far cheaper than any form of retraining. The limits lie in latency, in the quality of the retrieval component and in the fact that the model does not acquire real new skills. If you want to go deeper: RAG explains: How AI works with its own documents and Vector Database explains: Why RAG fails without good search are good entry points.

Fine tuning and LoRA are useful if the model is supposed to learn real new behaviors or styles, not just retrieve new factual knowledge. LoRA (Low-Rank Adaptation) drastically reduces the training effort by only training low-dimensional adapter matrices, not the entire network. However, catastrophic forgetting remains a risk, especially if the fine-tuning corpus differs greatly from the pretraining. Regularization can reduce this risk, but not completely eliminate it.

Model editing specifically changes factual associations in the weights. ROME is designed for single changes in GPT models. MEMIT extended this approach to thousands of associations in experiments with GPT-J and GPT-NeoX. This is a different type of problem than continuous learning from a stream of data. Whether changes outside of the tested facts remain specific and are stable over many update cycles must be checked separately.

Continuous pre-training continues pre-training with new data. This allows a model to accommodate broader knowledge, but data mixing, regression testing and computational effort remain challenging. For individual teams, this is significantly more complex than a RAG index or limited adapter training.

The Sleep Paper: Memory Consolidation as a New Approach

The preprint “Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories” by Ali Behrouz, Farnoosh Hashemi, Adel Javanmard and Vahab Mirrokni was published in June 2026. According to arXiv, an earlier version was public on OpenReview since September 2025. The biological sleep analogy frames a technical proposal with two phases.

Memory consolidation (knowledge seeding): Having learned new knowledge, one’s own smaller model distills its memories into a larger network via a combined distillation and RL process. This unusual upward direction is intended to create additional capacity while preserving existing knowledge.

Dreaming: In a second step, the model itself should improve. It creates a curriculum of synthetic data using reinforcement learning to rehearse new knowledge and refine existing skills without human supervision. The paper describes dreaming as a self-improvement phase, not primarily as a protective mechanism against forgetting. Our DeepDive organizes the larger context of such feedback AI Models That Evolve: The Recursive Revolution a.

The authors report, among other things, results with Qwen3-8B on AIME-24, AIME-25 and HMMT-25 as well as tests on knowledge absorption and long context. These are results of our own proof of concept. The preprint does not mention any independent replications. “Sleep” also requires training, additional model components and defined update cycles. The approach is a long way from a model that reliably updates itself during ongoing operation.

What this means for local setups

Anyone who runs a model like Llama 3 or Mistral locally, for example on a DGX Spark or comparable inference hardware, will encounter these limits in everyday life. The model has a knowledge cutoff. RAG is the first sensible answer to current factual knowledge, as long as the retrieval system is well built. Fine-tuning on small corpora is possible, but requires metrics on old benchmark tasks. Otherwise, a loss of performance will only be noticeable during operation.

Specifically: If you fine-tune a 7B model locally and then only test it on the new task set, you can see better results and at the same time miss a regression. Therefore, an unchanged base model, an old evaluation set, and a cutoff value for each protected capability are part of the experiment. DGX Spark: Local AI hardware in practice shows how such operating limits can be classified in local inference.

A customized model is no longer a quietly updated copy of its base model. It is a new artifact with its own data set, regressions and rollback requirement. For current knowledge, RAG usually remains the sensible operating decision today. Continual learning is research and not a maintenance-free update button.