Andrej Karpathy predicts 10 years for LLMs to achieve continual learning

Recently, we have reported on numerous startups and research findings focused on "continual learning," such as "Mind Lab Releases Latest LoRA Advances, Revealing a New Paradigm for Continual Learning in Large Models," "Breaking Free from KV Cache Constraints: Can Long Contexts Be Embedded into Weights to Enable Continual Learning in Large Models?" "ICML 2026 | Breaking Limits! HKU Introduces the First Continual Learning Architecture Adaptable to 300+ Tasks, Solving the Forgetting Problem," and "Can DeepSeek The Self-Evolving Harness! The LlamaFactory Team Opens Sources a New Tool: Automatically Create Agents for Just 0.2 Yuan. When “Getting Bigger” Is No Longer the Only Path, Another Chinese Model Is Open-Sourced…

Yes, it’s quite intense, and “continuous learning” is also becoming one of the most frequently encountered keywords. This reflects a repeatedly emphasized insight.

In October 2025, Andrej Karpathy said on Dwarkesh Patel’s podcast: “Current large models lack continuous learning. You can’t tell them something and expect them to remember.” He estimated it would take about another decade to address these cognitive limitations. Two months later, in his annual review, he placed this statement within a broader context: the major leap in model capabilities in 2025 came primarily from verifiable reward reinforcement learning (RLVR), but within the entire LLM stack, memory, multimodal perception, continuous learning, and the ability to operate computers remain clear weaknesses. “We have prototypes, but no agents ready to work as colleagues.”

Continual Learning (also known as lifelong learning) has become one of the hottest concepts over the past year.

It refers to the model’s ability, after deployment, to continuously learn from new tasks, new knowledge, and new experiences—without forgetting what it has already learned.

This sounds straightforward, but in practice, it’s extremely difficult—so difficult that it has become the toughest hurdle on the path to an “AI colleague.” The reports above also show that this journey is no longer a single path, but rather multiple diverging routes advancing simultaneously.

Over the past year, academia and industry have diverged into several distinct technical pathways in pursuit of enabling models to “learn while using.” Some approaches attach external memory to models, others continuously update weights, some opt for full re-pretraining, and a newer wave of ideas seeks to redefine what “learning” itself means. This article aims to lay out each of these directions clearly, explaining what each path is betting on and where it faces bottlenecks.

A core obstacle to continuous learning has a specific name: catastrophic forgetting. A neural network's knowledge is stored in billions of weights; when you fine-tune the model with new data to update these weights, the model often overwrites the parameters responsible for previous capabilities while learning new tasks, causing a sharp decline in performance on previously mastered tasks.

Diagram of catastrophic forgetting. When an artificial deep neural network is trained sequentially on two tasks, it rapidly and completely forgets the first task while training on the second.

This phenomenon was extensively studied in the era of small models, but it has introduced new challenges with LLMs. The TRACE benchmark, specifically designed to evaluate continual learning in LLMs, found that fine-tuning an already aligned model continuously not only causes it to forget previous tasks but also degrades its general capabilities and instruction-following ability. In other words, teaching the model something new may come at the cost of making it generally less capable and less responsive to instructions.

Precisely because directly modifying weights carries such high risk, "continual learning" has split into several schools of thought. Their fundamental disagreement lies in the different trade-offs made regarding whether to modify weights and where to store new knowledge.

The most straightforward and quickest approach is to leave the model weights unchanged and store new knowledge in an external database, retrieving it back into the context when needed. This path has evolved from RAG (Retrieval-Augmented Generation) and has now grown into a dedicated field: Agent Memory.

The representative work is MemGPT (the company behind it has since been renamed Letta). It draws an analogy between LLM context management and an operating system’s memory management, distinguishing between a limited “working context” and a larger “external storage,” enabling the model to autonomously decide what to bring into context and what to write back to storage, much like an OS manages memory.

Following this line of thought, a group of systems each have distinct focuses: Mem0 emphasizes production-grade, scalable long-term memory access; Zep enhances retrieval with a temporal knowledge graph to handle cross-session temporal reasoning; A-MEM draws inspiration from the Zettelkasten note-taking method, assigning structured tags to each memory and automatically linking them to relevant past entries, making retrieval more context-aware.

Karpathy himself is also betting on this direction. He repeatedly emphasizes that the future will not require "larger hard drives" of memory, but rather a compact "cognitive core" (which he estimates needs only one or two billion parameters) paired with a structured external memory system capable of self-reinforcing growth.

Following this idea, he released a model called "LLM Wiki" on GitHub: instead of fetching raw text chunks via RAG for every query, the agent proactively compiles the information into a continuously updated, interlinked knowledge base, which is then queried instead.

Karpathy's LLM Wiki documentation has garnered nearly 45,000 stars and has been forked over 9,000 times.

Letta themselves elevated this approach to the proposition of “continuous learning within the token space.” They pointed out that the current default practice is “append first, then summarize”—continuously stacking raw experiences until the context overflows, then compressing them into a summary. This has two flaws: appending shifts all representation work to inference time, requiring the model to reprocess raw logs during every forward pass; and summarization is lossy and abrupt, causing important details to vanish without warning. Letta’s bet is that memories learned within the token space will eventually be more valuable than the model’s own weights.

The advantage of this approach is its security, controllability, and interpretability—you can delete or correct errors anytime. The drawback is that it doesn’t truly internalize knowledge into the model; every response relies on retrieval through the narrow gateway of context. As the knowledge base grows, retrieval accuracy and cost become bottlenecks.

Taking one more step beyond external memory is a more sophisticated approach: keeping the weights unchanged while allowing the model’s input context to evolve continuously. This is known as context engineering.

In October 2025, ACE (Agentic Context Engineering), proposed by Stanford, SambaNova, and UC Berkeley, exemplifies this approach. It treats context as a continuously evolving "playbook," maintained by three specialized roles: the Generator creates reasoning trajectories, the Reflector extracts concrete insights from successes and failures, and the Curator organizes these insights into structured incremental updates, integrating them into the playbook.

ACE aims to address two common flaws in similar approaches: first, "brevity bias"—when LLMs repeatedly rewrite context, they tend to compress and discard domain-specific details; second, "context collapse"—repeated rewriting causes details to gradually disappear.

ACE avoids these two issues by using incremental delta updates instead of full rewrites. According to the paper, ACE achieves a 10.6% improvement over baselines on agent tasks and an 8.6% improvement on financial reasoning, while reducing adaptation latency by approximately 86.9%; on the AppWorld leaderboard, it achieves these results using smaller open-source models DeepSeek With ACE, V3.1 achieves an average score matching the production-grade agent IBM CUGA based on GPT-4.1.

Notably, ACE emphasizes that it can operate without labeled supervision; instead, it relies on naturally occurring feedback signals during execution—such as whether the code runs successfully or encounters errors—to guide reflection and refinement. This aligns it with the broader narrative of "agents learning from their own experiences."

External memory and context engineering avoid the risks of modifying weights, but they also bypass true knowledge internalization. Another group of researchers argues that, in the long term, certain knowledge and skills must ultimately be encoded into the parameters to be truly effective. This is known as continual post-training, which primarily consists of stages such as continual instruction tuning and continual preference alignment.

John Schulman of Thinking Machines offered a hierarchical perspective: he compared learning to psychological categories such as motor learning, episodic memory, and procedural memory, suggesting that in-context learning excels at handling short-term learning tasks, while parameter fine-tuning—including methods like LoRA—builds on top, particularly suited for tasks requiring greater capacity and true knowledge absorption; when the time horizon extends and in-context learning becomes insufficient, parameter fine-tuning prevails.

The biggest enemy of this approach remains catastrophic forgetting, and a recent interesting solution comes from Thinking Machines' Tinker.

Tinker is their first product, released in October 2025—a LoRA-based fine-tuning API that abstracts away the complexity of distributed training, exposing only low-level primitives like forward_backward and optim_step, allowing researchers to focus on data and algorithms.

On continuous learning, they champion a technique called Self-Distillation Fine-Tuning (SDFT): the core insight is that standard supervised fine-tuning is "off-policy," forcing the model to mimic tokens it would never generate itself, causing each new skill learned to erode previous capabilities; SDFT enables the model to act as its own teacher, learning new skills from demonstrations without forgetting old ones. A startup called Trajectory has already adopted Tinker as the core infrastructure for its continuous learning platform.

Incidentally, using LoRA instead of full fine-tuning offers a practical advantage: multiple fine-tuning tasks can share the same compute pool, spreading out the costs. This also explains why “continuous learning as a service” is becoming a business: the industry is currently exploring ways to turn frequent model updates into an on-demand API.

If fine-tuning is like making adjustments at the surface level of a model, then continual pre-training (CPT) goes back to the foundational layer, continuing to pre-train the model with new corpora so it adapts to new domains, languages, or evolving knowledge distributions over time. Compared to retraining from scratch by mixing new and old data together, CPT builds upon the existing model and is far more computationally efficient.

Its typical applications include knowledge drift over time, cross-lingual expansion, and cross-domain adaptation. However, the cost is equally clear: multiple empirical studies have consistently shown that continual pretraining is computationally expensive and prone to catastrophic forgetting of previously learned knowledge. As a result, the research community has been actively seeking continual pretraining methods that are free from replay and task annotation, aiming to achieve forgetting-free learning at the scale of LLMs.

For the vast majority of companies, the cost of retraining a state-of-the-art model from scratch is prohibitively high—this is precisely the origin of Karpathy’s statement that “large models are not suited for frequent retraining.” Therefore, true “retraining” is largely an option reserved for leading research labs, while continued pretraining offers a more practical compromise.

The first four directions still largely revolve within the binary framework of "cheating versus adjusting weights." However, a new wave of recent work has emerged over the past year, aiming to break free from this framework and redefine learning itself.

One approach is to have the model generate its own training data and decide how to update itself.

MIT’s SEAL (Self-Adapting Language Models) is a prime example. Given a new input, the model generates a segment of “self-edit”—a natural language instruction that specifies how to restructure the information, which hyperparameters to use for updating weights, and even which tools to invoke for data augmentation. The model then uses this instruction to fine-tune itself, resulting in persistent weight updates. The question of “what constitutes an effective self-edit” is trained by an outer reinforcement learning loop, with the reward signal being the model’s performance on downstream tasks after the update.

The NeurIPS 2025 version further demonstrates that this adaptive capability strengthens as model scale increases, and it mitigates forgetting through reinforcement learning. Its authors envision a model capable of autonomously deciding during inference whether to "learn something now," transforming one-time chain-of-thought reasoning into permanent ability.

Another more aggressive approach comes from Google’s Nested Learning, presented at NeurIPS 2025. Its core idea is to reinterpret a model as a set of nested, multi-level optimization problems, each with its own “context stream” and update frequency—under this perspective, architecture and optimization algorithms are unified as different levels of the same thing.

As a proof of concept, they developed Hope, a self-modifying architecture that extends the Titans long-term memory architecture with two key enhancements: unlimited nested learning layers and a "Continuous Memory System" (CMS). In simple terms, instead of a binary distinction between short-term and long-term memory, it features a full spectrum of memory modules that update across multiple time scales.

In experiments, Hope outperforms standard Transformers and modern recurrent models on language modeling, long-context reasoning, and continual learning tasks. Interestingly, subsequent work has introduced a "sleep" phase to this architecture—similar to how humans consolidate memories during sleep—allowing the model to allocate computational resources between active sessions to distill useful abstractions into more persistent parameter memory.

At a higher level of abstraction, it is the "Era of Experience" narrative proposed by David Silver and Richard Sutton.

Their assessment is that the knowledge extractable from human data in critical fields such as mathematics, code, and science is approaching its upper limit; to move forward, AI must continuously learn from its own experiences generated through interaction with the environment, forming a self-sustaining data loop. They view DeepMind’s AlphaProof in 2024—which won an IMO medal through “continuous interaction with formal proof systems”—as the beginning of this era. This narrative links continuous learning, reinforcement learning, and agent-driven autonomous exploration together, while also raising an unresolved question: who will design the reward functions that transform raw signals into useful guidance?

Another adjacent direction, often confused with continual learning but with a different objective, is knowledge editing. Represented by methods such as ROME and MEMIT, it achieves precise single or batch updates of specific facts by locating and modifying the MLP layers that store those facts, without requiring full retraining. However, its goal differs from that of continual learning: knowledge editing aims to precisely overwrite a specific fact, but in continuous, lifelong editing scenarios, repeated parameter modifications can interfere with each other and gradually accumulate “toxicity,” leading to model collapse. This has spurred the development of specialized methods such as WISE and AlphaEdit to address sequential editing.

When viewed together, these directions are arranged along an axis of "where knowledge resides."

The outermost layer consists of external memory and prompt engineering, where knowledge exists entirely outside the model in the token space—secure and controllable, but not truly internalized. The middle layer involves continual fine-tuning and continual pre-training, where knowledge is embedded into the model’s weights, offering higher capability limits but confronting the challenge of catastrophic forgetting. The innermost and most cutting-edge approach involves novel methods that enable the model to modify and evolve itself, aiming to make the very act of “learning” a core component of the model’s capabilities.

A pragmatic observation is that these approaches are likely not mutually exclusive but will instead collaborate in layers. Schulman’s hierarchical learning perspective and Karpathy’s concept of a “cognitive core + external memory” essentially convey the same idea: short-term, volatile knowledge is delegated to context and external memory, while long-term, deeply ingrained capabilities are entrusted to parameter fine-tuning—each playing its designated role. ACE’s ability to evolve through execution feedback without supervision, and SEAL’s use of reinforcement learning to optimize self-editing, both suggest a shared trend: future continuous learning will likely be driven by models themselves determining “what to learn and how to learn it.”

So, returning to Karpathy’s prediction of “another ten years,” has continual learning been solved? The answer is likely—no, but it’s no longer a blank slate. The core challenge of catastrophic forgetting has not yet been fully overcome; methods like SDFT and nested learning still only mitigate it rather than eliminate it. Can continual learning be solved solely through better context management and fine-tuning, or does it require entirely new ideas? This more fundamental question remains unresolved, even by Schulman himself.

It is clear that between 2025 and 2026, continuous learning transformed from a repeatedly mentioned "missing skill" into a vibrant, multi-path battlefield where startups are actively developing products. The path that first turns the "AI colleague" from prototype into reality is one worth continuing to watch.

Reference link

https://www.lesswrong.com/posts/qBsj6HswdmP6ahaGB/andrej-karpathy-on-llm-cognitive-deficits

https://karpathy.bearblog.dev/year-in-review-2025/

https://www.letta.com/blog/continual-learning/

https://arxiv.org/abs/2510.04618

https://arxiv.org/abs/2402.01364

https://thinkingmachines.ai/tinker/

https://arxiv.org/pdf/2604.05096

https://arxiv.org/pdf/2606.03979

This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by Machine Heart's Continuous Learner, edited by Panda.