A bounded context window retrieves relevant episodes, preserves rules, consolidates topics, and selectively writes to long-term memory—without replaying the entire conversation to the model.