Why Does AI Forget What You Said? Context Windows and the Truth About Memory

You told the AI assistant your budget was $600, but it recommended a $900 option. This fictional example reveals a key distinction: the chat history you can scroll back through is not necessarily all the information the AI actually uses when generating its next reply.

Three Easily Confused Concepts

According to an analysis by technical writer Ahmadreza Samadi, understanding this issue requires first distinguishing three layers:

• Context: The information available for this reply, including your question, instructions, message history, and relevant file or tool results.
• Memory: Information that has been selectively saved, provided the application supports this feature.
• Training: The process by which the model learns underlying patterns, unrelated to any single conversation.

The blurred boundaries between these three are the root cause of user confusion like "the AI clearly saw what I said."

The Context Window: A Limited Workbench

The context window is a fixed amount of workspace available to the model when generating a reply, and its size depends on the specific model. According to Anthropic's official documentation, text, images, and newly generated responses all consume this space.

Think of it as a workbench: the surface area is limited, and materials that don't fit cannot be used in the moment. This is just an analogy, not a literal description of the storage mechanism.

Context Window is measured in tokens, not words or characters. A token is the smallest unit of text the model processes. In English, roughly 4 characters or 0.75 words correspond to 1 token, while in Chinese, one character typically takes about 1 to 2 tokens. So a window nominally sized at 200,000 tokens can actually hold far less Chinese content than intuition would suggest.

Window size is determined by model architecture and training configuration — it's an inherent property of the model, and users cannot temporarily expand it mid-conversation. When input approaches the limit, the system must make trade-offs: truncating the earliest conversation, compressing middle content, or triggering summarization. This is why early information in long conversations tends to "evaporate" — it's often the first part to be sacrificed.

How Chat Applications Manage Long Conversations

A chat application may summarize early conversation, omit some content, or retrieve historical information when needed. According to Anthropic's context engineering documentation and Claude's chat search and memory feature description, different products handle this in different ways.

This means: the complete chat record you see and the input the model actually receives may not be exactly the same. The application interface displays the transcript, while the model consumes a filtered or compressed version.

Why Summaries Lose Key Details

According to Anthropic's analysis of effective context engineering, a shorter summary may preserve the main goal while missing precise constraints, dates, or decisions.

Back to the fictional example from the beginning: the core intent of "plan a trip" survives in the summary, but the hard constraint of "budget of $600" disappears. Summarization itself helps the conversation continue, but what is kept and what is discarded directly determines the quality of subsequent replies.

This phenomenon is cross-verified across multiple sources — Anthropic's context window documentation, context engineering guide, and product help docs all point to the same conclusion: the compression process is lossy.

Summarization is one implementation of lossy compression. Unlike lossless compression, which can fully reconstruct the original text, lossy compression aims to preserve the "semantic backbone" within limited space, at the cost of discarding details deemed secondary.

The problem is that the criterion for "secondary" is determined by the model or system rules, not by the user. Models tend to preserve topics, intent, and emotional tone, since these matter most for generating coherent replies; whereas numbers, dates, negations (such as "don't" or "except"), and precise conditions are often treated as inferable or ignorable information. Yet for users, it is precisely these details that constitute hard constraints. Understanding this bias can help you proactively repeat important constraints in critical tasks, rather than relying on the system to preserve them automatically.

Available Does Not Mean Correctly Used

A subtler problem: even if a detail is still in the model's input, the model may ignore or misuse it. A larger context window does not guarantee better recall or correct answers.

Here we need to distinguish two different failure modes:

1. Loss of access — The constraint no longer appears in the input due to summarization or truncation.
2. Failure to follow — The constraint is still in the input, but the model does not use it correctly.

According to Anthropic's documentation, the strategies for these two situations are completely different. The former requires re-supplying the information, while the latter involves limitations in the model's reasoning ability.

The Difference Between Saved Memory and Current Context

Some applications support saving selected information across conversations or retrieving relevant earlier chats. According to Claude's chat search and memory feature description, these features depend on product design, user settings, and access permissions, and are not equivalent to "stuffing every old message into every reply."

A practical tip: regularly review important saved notes, rather than assuming the system always maintains a complete, permanent transcript of the conversation.

Memory features are typically implemented technically through vector databases or structured storage: the system writes selected information into persistent storage, retrieves it by relevance when needed, and injects it into the current context. The key difference from the context window is lifecycle — memory persists across conversations, while context is only valid within a single request.

But retrieval relies on similarity matching, which may recall irrelevant information or miss content with similar semantics but different wording. In addition, writing to memory usually requires explicit user action or system judgment, and not all conversation content is automatically saved. Therefore, "I said that last time" does not automatically hold true in a memory system unless that information was actually saved and successfully retrieved.

Sending a Message Is Not Training the Model

According to IBM's explanation of AI inference, training adjusts the model to learn patterns; whereas everyday responses (called inference) simply use the trained model to process current input. Sending a chat message does not immediately retrain the model.

However, service providers may have separate policies regarding the subsequent use of chat data. It's advisable to check specific settings and terms to understand how your data is handled.

Inference and training are two distinct stages in the machine learning pipeline. Training is the process of repeatedly adjusting model parameters (weights) through large amounts of data — it takes weeks or even months, is extremely costly, and is usually done offline. Inference uses already-trained fixed parameters to process new input and generate output, and the parameters do not change during inference.

This means that every conversation you have with the model, by default, does not "teach" it anything — the model won't permanently remember just because you corrected it once. So-called "the model got smarter" usually means the vendor released a new version, not that your conversation had an effect. Unless a product explicitly provides and enables feedback-based fine-tuning or memory writing, a single conversation has no lasting impact on the model itself.

Practical Strategy: Maintain a Short, Up-to-Date Brief

For long-running tasks, Ahmadreza Samadi recommends restating goals, hard constraints, and the latest decisions together. Specific practices include:

• Asking the assistant to list these constraints before answering
• Comparing the results against the constraints point by point
• Verifying important names, dates, and numbers in the final output

A fresh brief can provide effective input, but cannot guarantee correctness. This approach also corresponds to recommendations in Anthropic's context engineering guide.

Key Takeaways

AI "forgetting" what you said usually isn't caused by a single reason. It could be that summarization lost details, that the context window was full, or that the information was still there but the model didn't use it correctly. Understanding the differences between context, memory, and training can help you set more reasonable expectations when collaborating with AI and take more targeted countermeasures.

One-line Summary

AI "forgetting" what you said stems from confusion between context windows, memory, and training.

Paragraph Summary

The article points out that users are often puzzled when AI clearly saw some information but didn't use it, and the root cause is confusing context, memory, and training. Context is the limited input space available for a single model reply, constrained by the context window size; memory is selectively saved cross-conversation information; training is unrelated to any single conversation. Chat applications may summarize or selectively retrieve historical conversations, causing the record users see to differ from the input the model actually receives. Summary compression is lossy and may preserve main goals while losing precise constraints. The article also distinguishes between "loss of access" and "failure to use correctly," and emphasizes that sending a message is not training the model. A practical suggestion is to proactively maintain a short brief containing goals, hard constraints, and the latest decisions.