A context window is the working memory of an AI model for a single conversation. Everything the model can take into account at that moment has to fit inside it: the system instructions, your messages, its earlier replies, any files or error logs it has read, and the answer it is writing. The size is measured in tokens, not words or characters.
Context windows have grown large, but they are still finite, and coding work fills them fast. A few big files, a long error trace and an hour of back-and-forth add up. When the window is full, tools handle it in different ways. Some drop or summarize older parts of the conversation. Claude Code, for example, can compact a long conversation into a summary, either automatically or when you run /compact.
The practical effect is that a long session can forget details from the start, or give them less weight. If the model starts repeating an earlier mistake or ignoring an instruction you gave an hour ago, a full context is a likely reason.
Habits that help. Start a fresh conversation for each separate task. Put rules that should always apply, like your stack, your style and what not to touch, in a project instructions file such as CLAUDE.md or your editor’s rules file, so you do not repeat them. Point the model at the specific files that matter instead of the whole project. Paste the relevant part of an error, not the full log.
More context is not always better. The most useful context is the smallest amount that still contains the answer.
In the wild
This chat is getting long and it’s forgotten we’re on Postgres. Starting a new session with just the schema and the failing test.
