Why Your AI Forgets Everything (Even Though It Feels Smart)

Welcome to Decoding AI with Tarrush

Artificial intelligence is becoming part of almost everything we do, yet most people only see what happens on the surface. We interact with chatbots, generate images, write code, and search for answers, but very little is said about what is actually happening behind the scenes.

That is what this series is about.

In Decoding AI with Tarrush, we will explore the ideas, challenges, breakthroughs, and unanswered questions that are shaping modern AI. Every article focuses on one concept and explains it in a way that is easy to understand without losing the science behind it.

This is not about chasing headlines or making bold predictions. It is about understanding how AI really works, where it struggles, why researchers are working on certain problems, and what those developments mean for everyone who uses this technology.

Whether you are a student, a builder, a professional, or simply curious about AI, I hope each article leaves you with a clearer understanding than when you started.

Let’s begin.

Photo by Steve A Johnson on Unsplash

You spend an hour teaching an AI chatbot your preferences, correcting its mistakes, maybe even training it on how to do your work. Then you close the tab. The next day, poof it remembers nothing. It’s not because the app is broken or the company is lazy. Instead, this is due to catastrophic forgetting, an inherent limitation of neural networks first identified in 1989.

In short, AI models today start each conversation like a blank slate. They do not accumulate memories across sessions the way humans do, so anything you “taught” them is effectively wiped clean once the session ends.

Why AI Can’t Just “Remember and Keep Learning”

Consider a classic experiment: a human learner memorizes list A →B pairs and then learns a new list A → C. Humans retain much of the first list even after learning the second. In contrast, a typical neural network sees its performance on the first task (red curve) drop to near zero once it learns the second. This happens because the AI’s “brain” is just a fixed set of weights that gets updated during training. When new data arrive, the model tweaks those weights often overwriting the configurations that encoded the old knowledge.

In other words, if the model is sensitive enough to learn the new task, it tends to destroy the old representations. As one analysis puts it, large language models “don’t remember you” each conversation is treated independently. Every chat session starts from the same factory-reset state.

Catastrophic forgetting was first noted by McCloskey and Cohen in 1989, and after nearly four decades it remains unsolved. In fact, recent research shows the problem often worsens as models grow. A study of modern LLMs (1–7 billion parameters) found that larger models actually suffered more forgetting than smaller ones. In other words, simply scaling up an AI today doesn’t magically fix memory; it can exacerbate the issue. Until we invent a new learning paradigm, teaching an AI something new will usually mean risking what it learned before like painting over the wall with thicker paint until the old layer vanishes.

Why This Matters Right Now

For most of AI’s history, this forgetting was hidden. Companies would train a model and then deploy it without further updates, so the model was effectively “frozen.” Users didn’t expect it to adapt because it never changed. Today, expectations have shifted. People now want their AI assistants to be personal: remembering their name, preferences, past projects, writing style, and more. As one recent industry commentary notes, users increasingly want AI to retain “preferences, past work, writing style, and business information across tasks,” providing continuity from one chat to the next.

In other words, the gap between what models currently know and what they could know has become glaringly obvious. Chatbots like ChatGPT, Claude, Gemini and many startups are racing to add “memory” features or tools that simulate it, because customers demand that their AI feels like it learns and grows over time.

The core challenge is that achieving true memory in AI requires solving several hard problems at once:

• No easy replay: Neural networks normally learn by training on batches of data. In practice, we can’t keep an ongoing archive of everything a model has ever seen or been taught. Storing all past inputs or training data is impractical (and often prohibited) for privacy and scale reasons. Without a perfect replay buffer, the model can’t simply re-expose itself to old data when learning new.
• No “new info” flag: Human learners often recognize when they’re learning something new, but AI models aren’t given such cues. When presented with new text or tasks, a model has no built-in signal saying, “be careful this updates your knowledge.” It treats every example uniformly, so new learning can inadvertently override old skills.
• Must work at scale: Solutions that work in small academic demos often break when applied to a deployed LLM with billions of parameters. Studies find that larger models actually experience more forgetting, and nothing has been proven at the scale of today’s AI systems. Any fix has to handle the massive size and speed of commercial AI services.
• Automatic, zero-touch: In deployed products, memory updates must occur without humans babysitting. We cannot rely on engineers manually deciding which memories to save or how to retrain the model after every chat. True continual learning would need to happen autonomously (or at least with minimal human oversight), which is an unsolved engineering problem.

No system today checks all these boxes. For instance, even if we could magically archive every conversation, we’d face huge privacy and management headaches. And if we let the model update itself freely, we’d need to safely orchestrate those updates at real-time scale without bugs or biases creeping in.

What People Are Actually Doing About It

Instead of cracking the impossible directly, engineers have mostly resorted to clever workarounds:

• External memory modules: Some systems give the AI a separate “notebook.” Rather than changing the main model weights, the AI writes down important facts or context into an external store (a vector database, knowledge graph, etc.). At each turn, the model can read relevant notes and use them as extra context. These memory-augmented networks act like having unlimited post-it notes attached.
• For example, techniques like Gradient Episodic Memory (GEM) allow a model to save and recall past examples explicitly, preventing them from being overwritten. In practice, companies like Pinecone or Memworks provide these memory backbones so that the core AI stays frozen but can look up past conversation pieces on demand.
• Session-specific fine-tuning: Other teams let the model tweak itself briefly during a chat, then throw those tweaks away afterward. In this approach, the AI “learns” from the current user’s inputs on the fly but then resets to its original state after the conversation. It’s like giving the model a temporary costume for that interaction. This can improve performance on the spot without committing any permanent changes (and thus nothing “old” is lost). In machine learning terms, it’s similar to test-time training: the model runs gradient descent on the immediate task and then discards that update. Nothing inside the model is changed permanently, so by design there’s nothing to forget later.
• Pure retrieval/in-context tricks: A very popular modern approach is to essentially avoid internal learning altogether. Here, all memory lives outside the model, and the system simply fetches or re-injects it as needed. Many products rely on large context windows (Just like we see in the claude code), chain-of-thought prompts, or retrieval-augmented generation (RAG) to give the illusion of memory. With RAG, for instance, the model’s knowledge is augmented by an external database of documents. Every time the AI needs to answer, it searches that database and includes the relevant facts in the prompt, instead of having those facts baked into its weights. Since the base model never actually integrates new information into its parameters, it never truly “forgets” (because nothing ever changed). This is effectively a more sophisticated form of copy-pasting knowledge into each chat.

Each of these hacks can help in practice, but none is a panacea. External memories can get cluttered or inconsistent over many chats. Session-only updates don’t carry any learning forward to the next day. Pure retrieval means the AI never internalizes new knowledge in a deep way. As one analysis starkly puts it, today’s solutions are all “scaffolding” around a fundamentally static model. The AI may accumulate context on the surface, but under the hood it hasn’t changed: the model “itself never fully internalizes the new information”.

In other words, behind the scenes these are still temporary fixes, not the long-term learning breakthroughs we really want.

The Simple Takeaway

In practical terms, today’s AI is phenomenal at regurgitating what it was originally trained on, but it’s still very poor at safely learning new, persistent knowledge from your interactions. So, when a product touts an AI that will “remember everything about you and grow smarter,” take it with a grain of salt. Usually it just means they’ve implemented one of the above tricks a bigger context window, a private memory store, a fancy prompt, etc. not that the AI has suddenly become self-aware. Knowing about catastrophic forgetting helps you see which parts of an AI’s behavior are genuine progress and which are just clever engineering hacks.

In short: be impressed by what AI can do, but stay skeptical of claims that it’s truly learning like a human when it seems to forget everything once you close the chat.

Up next in this series: Article 2 will dive inside the AI’s “mind” how researchers are trying to interpret its reasoning and why understanding that process might be the next big step in making AI smarter and more reliable.

Sources: Our explanation here draws on AI research and industry commentary (e.g. reviews of continual learning and AI memory features) to reflect the current state of the field.

Thanks for Reading

Follow for more updates on LinkedIn — Tarrush Saxena | LinkedIn