The set of architectural and algorithmic techniques used to extend the effective memory of large language models beyond their native token limits, enabling coherent processing of long-horizon inputs.
Overview
- OpenAI’s GPT-5.1 Codex Max is the first model natively trained to operate across multiple context windows through a process called compaction, enabling coherent work over millions of tokens in a single task. (Source: OpenAI, via AI Daily Brief, 2026-08-24)