A development team is building a language model intended to act as a legal assistant, primarily tasked with summarizing lengthy court transcripts and contracts that can be hundreds of thousands of tokens long. They are facing significant memory and computational bottlenecks during inference due to the size of the context window. The team is debating two architectural approaches to solve this problem:

1. Implementing a modified attention mechanism where each new token only attends to a fixed-size window of recent tokens and a selection of globally important tokens from the distant past.
2. Integrating an external, fixed-size memory state that is updated after every block of tokens, compressing the information from that block into the memory before it is discarded.

Evaluate the potential trade-offs of each approach for this specific legal summarization task. In your evaluation, consider aspects like information fidelity (risk of losing critical details), computational efficiency, and implementation complexity.