Think of it as the boundary around the AI’s working memory. Everything the model holds at once (instructions, the conversation, file content, tool output) has to fit inside, and under the hood that limit is measured in tokens.
We explain it with an employee in a lecture series. Drop you into seven lectures in one day and test you at the end, and you’d remember the main idea but miss the detail from hour three. Overload isn’t obvious right away either. The student still sounds confident on the easy questions, and the misses show up when the work depends on a specific earlier detail.
How it shows up
In a long session, the AI may forget earlier details, mix instructions, or spend effort just holding the thread together. That doesn’t mean the model got worse; often the session is just crowded. That’s why long work gets split into phases, and why subagents help: one researches a narrow question in its own smaller context, then reports back, so the main session doesn’t carry every detail. When the window gets too full, some tools use compaction to summarize so work can continue. That helps, but it’s lossy, like your notes from a long meeting aren’t the meeting itself.
Why you care
Don’t memorize exact numbers; they change by model, product, and plan. The steadier lesson: keep sessions focused, split big work into phases, save durable context into files, and wrap up before the room gets crowded. Even a capable model can only work with what still fits in the room.