When people first start using ChatGPT, there's often confusion about what it "remembers." Does it learn from our conversations? Does it know what we talked about yesterday? The answer is more nuanced than yes or no, and understanding it makes you a more effective user.

The Mental Model Problem

Humans have persistent memory. We remember yesterday's conversations, last week's meetings, childhood experiences. When we talk to ChatGPT, we instinctively expect the same.

But ChatGPT doesn't have memory in this sense. It has a context window—a fixed-size buffer that holds the current conversation. Everything the model "knows" about your interaction exists within this window. When it fills up, older content gets pushed out.

What This Means Practically

In a long conversation, ChatGPT will eventually "forget" what you discussed at the beginning. This isn't a bug—it's a fundamental constraint of the architecture.

Signs you've hit context limits:

  • The model contradicts something it said earlier
  • It asks for information you already provided
  • Responses become less coherent with prior context
  • It "forgets" established conventions or formats

Working Within the Constraints

Once you understand the model, you can work with it rather than against it:

Strategic Summarization

Periodically ask the model to summarize the key points of your conversation. Then start a new conversation with that summary as context. You're essentially doing manual memory management.

Explicit Context Injection

Don't assume the model remembers. If you're continuing a previous line of work, re-state the relevant context at the start. "We're working on X. The current state is Y. The goal is Z."

Chunked Conversations

Break large tasks into smaller, self-contained conversations. Each conversation has a focused goal and doesn't rely on accumulated context from hours of prior work.

The Bigger Picture

Understanding context windows isn't just about ChatGPT. It's a window into how these systems work. LLMs don't "think" between conversations. They don't "learn" from your interactions (unless explicitly fine-tuned). Each conversation starts fresh.

This has implications for trust, privacy, and how we should design AI-assisted workflows. The model isn't building a profile of you. It's not secretly remembering sensitive information. But it's also not learning your preferences over time.

Note (2026): This was written in 2023. Since then, persistent memory features have been added to ChatGPT. But the core concepts about context windows remain relevant—memory features are built on top of this architecture, not replacements for it.

Understanding these constraints shaped how I approached my music theory experiments with ChatGPT. Knowing the model wouldn't "remember" made me more deliberate about structuring conversations and extracting useful outputs before context was lost.