Compact Attention: Training-Free 1M+ Context with Perfect Recall

DCA gave us training-free context extension in 2024. Two years later, every assumption it made is outdated. Compact Attention rebuilds the idea from scratch for 2026 hardware, retrieval heads, and agent workloads.

Tejas Bhakta
Tejas Bhakta
February 27, 202613 min read
Compact Attention: Training-Free 1M+ Context with Perfect Recall