I used this for a bit and I felt like it was slower and generally worse than using 200K with context compaction. Context compaction does lose some things though.