NEW DELHI, Sept. 22
Artificial intelligence systems are increasingly expected to remember preferences, instructions and information across conversations, but users can still find an AI apparently forgetting something they told it minutes, days or even months earlier. The problem, sometimes described as “AI amnesia”, has less in common with human forgetfulness than the term suggests.
What appears to be memory loss can arise from several different parts of an AI system. Information may fall outside the context available during a conversation, become difficult for the model to identify among large amounts of material, fail to be retrieved from an application’s memory system or behave differently after the underlying model is updated.
Understanding those distinctions is becoming more important as AI develops beyond standalone chatbots into assistants and agents expected to work with users over long periods and across documents, emails, databases and other sources.
A language model does not remember a conversation in the same way a person remembers an experience. When generating an answer, it works with information made available to it at that particular moment.
Earlier messages can form part of that information, known as the model’s context. But in a basic AI system, the model does not independently reach back into previous conversations and retrieve everything a user has said. The application surrounding the model has to make relevant information available again.
This means what users experience as AI memory is often a combination of several systems rather than a single store of information. An application can save user preferences, summaries or other information in databases and retrieve them when they become relevant to a later conversation.
OpenAI, for example, describes ChatGPT’s memory system as a way of carrying useful information and preferences from previous conversations into later ones. The underlying idea is straightforward: information stored outside the immediate exchange can be retrieved and supplied to the model as additional context.
Problems can arise even within a single conversation because every language model has limits on how much information it can process effectively at one time.
A model’s context can include the current question, earlier messages, instructions, uploaded documents and information supplied by other systems. As that material grows, identifying which details matter for a particular answer becomes increasingly difficult.
A large context window does not necessarily mean every piece of information inside it receives equal attention.
Experiments cited by Business Standard found that models could become less consistent as more information was added, even before reaching their stated context limits. The behaviour has been described as “context rot”. Information buried in the middle of a large amount of material can be overlooked or confused with details appearing elsewhere.
In such cases, the AI may not have erased anything. The relevant information can still exist within the material provided to the model but fail to influence the answer strongly enough.
That distinction separates context from memory.
Context is the information available to the model while it generates a particular response. Memory refers to information an AI application stores or retrieves so that it can potentially be used again later.
If a user tells an AI assistant how they prefer something to be written, for example, an application could store that preference. During a future conversation, the system might retrieve the preference and place it into the context supplied to the model.
The result can look like human memory from the user’s perspective, even though the underlying mechanism is quite different.
There is another form of forgetting that occurs deeper inside AI models.
Researchers call it “catastrophic forgetting”. This can happen when a model is trained on new information or fine-tuned for a new task and its performance on something it previously learned deteriorates.
IBM describes the problem as new training interfering with knowledge associated with earlier tasks. Because training changes the model’s internal parameters, improvements in one area can sometimes affect capabilities elsewhere. Researchers have developed techniques including rehearsal and regularisation to reduce the problem.
Catastrophic forgetting is fundamentally different from an AI assistant failing to remember something a user said in a previous conversation. The former involves changes to the trained model itself. The latter can result simply from information not being retrieved or supplied when the next answer is generated.
Model updates add another layer of complexity. AI systems are regularly retrained, fine-tuned and modified. An updated model may therefore respond differently to information that an earlier version handled reliably. From the user’s perspective, the change can look like something has been forgotten when what has actually changed is the model’s behaviour.
The challenge becomes greater as AI systems are given access to increasingly large quantities of information.
An AI agent working across thousands of emails, documents, previous conversations and databases might technically have access to everything it needs. But access alone does not guarantee that the correct information will be retrieved when required.
Models can miss relevant material, confuse details from different documents or produce inconsistent summaries as the amount of available information grows. Such failures can occur even when the total material remains within the model’s technical context limit.
Researchers are also studying the opposite problem: how to make AI deliberately forget.
Known as machine unlearning, the field examines whether particular information can be removed from a trained model without retraining it entirely. Potential applications include removing private, copyrighted or outdated information.
That is considerably harder than deleting a conventional computer file. Knowledge inside a trained model is distributed across patterns and relationships within its parameters rather than stored as individual documents in clearly identifiable locations. Removing one set of information can therefore affect unrelated capabilities.
Researchers must also establish whether information has genuinely disappeared. Something a model appears unable to reproduce under ordinary questioning could potentially re-emerge when prompted differently.
For casual chatbot use, these distinctions may seem largely technical. If an assistant forgets a minor preference, the result may amount to little more than inconvenience.
The consequences become more serious when AI systems are expected to manage long-running projects, work with business information or operate as semi-autonomous agents. An assistant that loses track of an important instruction, retrieves an outdated preference or confuses information from different documents can produce an answer that appears coherent while relying on the wrong context.
The central challenge for AI developers is therefore no longer simply giving models access to more information. Systems must decide what information deserves to be remembered, retrieve it when it becomes relevant, recognise when it has become outdated and prevent older information from being confused with newer instructions.
As AI assistants take on longer and more complicated tasks, their usefulness may increasingly depend not just on how much they know, but on whether they can reliably find the right memory at the moment it matters.
