AI Session

An AI session is the full conversation you have with a model from the first message you send to the moment you close the thread or start a new one. Everything typed, every reply, every code block, every revised draft lives inside that thread. The model reads the entire conversation history on every single response. That is how it knows what you said three exchanges ago. That is also why sessions have a size limit.

The concept exists because large language models are stateless by design. There is no background process running on your behalf between conversations, no sleeping agent holding your context. Every session is a fresh instantiation of the model, handed the full conversation history and asked to respond. When the session ends, nothing persists. You are starting from silence next time.

A session is not your account, and it is not your history. When you log into ChatGPT and see a left sidebar full of past conversations, those are transcripts, not active context. The model does not carry any understanding from conversation 47 into conversation 48 unless you paste something in. Seeing your old chat listed in the sidebar is like finding a journal entry from someone else. You can read it. The model in front of you right now cannot.

It is also not the same as a context window, though the two terms get conflated constantly. The context window is the maximum amount of text a session can hold before the model begins dropping the earliest parts. The session is the conversation itself. The context window is the ceiling. Confusing them is like confusing a road trip with the size of the gas tank: connected, yes; identical, no. You can have a short session inside a massive context window, or you can push a session far enough that it bumps the ceiling and starts forgetting its own earlier output.

Features like ChatGPT's Memory, which OpenAI introduced in early 2024 and rolled out to Plus subscribers that spring, do not eliminate sessions. Memory writes summaries to a persistent store outside the session. Those summaries get injected at the start of future ones. The session boundary still exists. Something is just taking notes for the next stranger you meet.

Cursor, the AI-native code editor that reshaped how developers interact with AI in 2024, builds its Composer feature around sessions. Every new Composer thread is a new session with whichever model you have selected: Claude Sonnet, GPT-4o, or Gemini. The model in that thread has no idea what you were debugging in the previous thread, what architectural decision you made yesterday, or why you named the function that way. Developers who treat each Composer thread as disposable constantly re-explain context. Developers who manage sessions deliberately, keeping all the relevant files and decisions inside one open thread, get noticeably better output over the course of a working day.

Anthropic's Claude 3 family, released in March 2024, ships with context windows up to 200,000 tokens on the Opus tier. That is roughly 500 pages of text. A single session can hold a full technical spec, a meaningful slice of a codebase, and a long conversation about both without dropping anything. That sounds like a lot. For a focused sprint on one problem, it is. For a project that spans weeks of real work, it evaporates fast.

Sessions work in your favor when the work is bounded and continuous. A focused design critique, a single article draft, a code review for one pull request: these fit cleanly inside a session. Keep the thread open, keep feeding it context, and the model compounds its understanding. Output at the tail end of a rich session is almost always sharper than output at the start, because the model has accumulated constraints, preferences, and concrete examples to work against.

Sessions work against you when you expect continuity that is not structurally possible. Closing Claude on Tuesday and reopening Wednesday expecting it to remember your client's brand voice, your preferred sentence length, your rejected headline ideas, and the three constraints your art director added last week is a category error. None of that survived the boundary. This is why system prompts exist. This is why teams that work seriously with AI build habits around what they reintroduce at the top of every session, because the model will not do it for them.

A session is not a relationship; it is a conversation with a stranger who forgets you the moment you leave the room.