Large language models (LLMs) today operate like someone with severe amnesia, frozen in knowledge from training but unable to learn from new experiences. They can access vast information through context windows and retrieval systems, but they cannot update their core understanding based on what they encounter after deployment. As AI agents tackle increasingly complex, multi-step tasks, this limitation is becoming a critical bottleneck.

The challenge is straightforward: in-context learning (ICL), which feeds information into a model's input window, works well for problems where answers already exist somewhere in the world. But for tasks requiring genuine discovery, adversarial scenarios like security threats, or knowledge too nuanced to express in language, models need a way to compress and internalize new information directly into their parameters after they're deployed.

Why Are AI Agents Failing After 20 to 100 Steps?

As AI systems shift from simple question-answering to autonomous agents that operate in loops, pressure is mounting on the in-context learning model. A single task can now consume a significant portion of a model's available context window. Each step in an agent's loop relies on context passed from previous iterations, and coherence degrades as the context fills up. The result: agents often fail after 20 to 100 steps because they lose the thread of what they were doing.

This is why major AI labs are investing heavily in models with very large context windows. The most common approach uses state space models (SSMs) and linear attention variants, which intersperse fixed memory layers with normal attention heads. These architectures offer better scaling for long contexts than traditional attention mechanisms. The goal is to help agents maintain coherence over longer loops, potentially from around 20 steps to 20,000 steps, without losing the breadth of skills and knowledge that transformers provide.

However, even this approach has limits. External memory layers and expanded context windows are non-parametric solutions, meaning they don't update the model's core weights. Some researchers argue this is not enough.

What Is Continual Learning and Why Does It Matter?

Continual learning refers to the ability of AI models to update their parameters in response to new experiences after deployment. Unlike in-context learning, which is transient and temporary, continual learning compresses new information directly into the model's weights. The concept is not new in AI research, but it has become increasingly urgent as the gap between what models know and what they could know has widened.

The analogy is human development. A 15-year-old prodigy might have exceptional foundational skills and knowledge, but they lack the lived experience and tacit understanding that comes from years of deployment in the real world. Humans learn continuously throughout their lives, updating their understanding based on new situations. Current AI models, by contrast, are frozen at deployment.

"The thing that happened with AGI and pre-training is that in some sense they overshot the target. A human being is not an AGI. Yes, there is definitely a foundation of skills, but a human being lacks a huge amount of knowledge. Instead, we rely on continual learning. If I produce a super intelligent 15-year-old, they don't know very much at all. A great student, very eager. You can say, 'Go and be a programmer. Go and be a doctor.' The deployment itself will involve some kind of a learning, trial-and-error period. It's a process, not dropping the finished thing," stated Ilya Sutskever.

Ilya Sutskever, Researcher

How to Implement Continual Learning in AI Systems

• Parametric Updates: Allow models to modify their weights based on new experiences after deployment, enabling genuine compression of new knowledge rather than relying solely on context windows or external retrieval systems.
• Memory Architecture Learning: Train models to develop their own memory structures rather than offloading memory management to external harnesses, potentially unlocking new dimensions of scaling and capability.
• Hybrid Approaches: Combine non-parametric solutions like expanded context windows with parametric learning mechanisms, using state space models to maintain coherence while enabling weight updates over time.

The Case for In-Context Learning Still Works, For Now

It's important to acknowledge that in-context learning is genuinely powerful and has driven remarkable progress. Transformers are conditional next-token predictors over sequences, and they can exhibit surprisingly rich behavior without touching their weights. Context management, prompt engineering, instruction tuning, and few-shot examples have all proven effective because the intelligence lives in static parameters, and apparent capabilities change radically depending on what is fed into the input window.

Recent examples demonstrate this. Cursor's autonomous coding agents achieve much of their capability through careful prompt orchestration, not special model access. OpenClaw, another agent system, broke out not because of exclusive model access but because it effectively turns context and tools into working state, tracking actions, structuring artifacts, and maintaining persistent memory of prior work.

The reason these non-parametric approaches have succeeded is that they are native to the transformer architecture, require no retraining, and scale automatically with model improvements. As models get better, prompting gets better. This "janky but native" interface couples directly to the underlying system rather than fighting it, which is why it has won so far.

Where Does Continual Learning Go From Here?

The question is not whether today's context-based systems work. They do. The question is whether we are looking at the ceiling of what in-context learning can achieve, and whether new approaches can take us further.

Researchers working on continual learning believe the answer is yes. As agents become more autonomous and operate over longer horizons, the limitations of fixed parameters will become increasingly apparent. The ability to learn continuously, to compress new experiences into model weights, and to develop internal memory architectures could represent a fundamental shift in how AI systems scale and improve over time.

This is why continual learning is considered some of the most important work happening in AI right now. The field sits at the intersection of scaling, deployment, and genuine capability advancement, offering a path beyond the perpetual present that current models inhabit.