Context overflow
Short answer
Context overflow is what happens when the prompt, the retrieved passages and the conversation together exceed the model's window — the oldest material is dropped.
The failure is quiet. The model does not announce that it has forgotten the start of the conversation; it simply answers as though it never happened, which reads like carelessness rather than a limit.
Starting a new conversation for a new document is the simple fix, and on a long document it is worth asking narrower questions so retrieval has less to fit.
Related terms
Other ways people ask this
- what is context overflow