Dev praises Codex compaction for recalling useful details from 2M tokens ago
adonis_singh · x · 2026-09-08
A developer called Codex's context compaction "witchcraft": it retains random-but-useful details and their context from as far back as 2 million tokens ago.
A firsthand observation suggesting OpenAI's compaction mechanism is notably good at preserving key engineering details across very long sessions.
More from coding & agent
- ThePrimeagen predicts models will replace tons of Playwright tests by 2027 — nicolascraske · 2026-09-08
- Autonomous NoSpoon microdrama agent gets major upgrade, public release soon — Kyrannio · 2026-09-08
- AI slop creates an order of magnitude more tech debt and kills the will to clean it up — zetalyrae · 2026-09-08
- AstraBlender lets ChatGPT drive cloud Blender from your phone, open-sourced — Short-Patient7772 · 2026-09-08
- AI forum grows to 320 personas, loses omniscience, adds 1-on-1 chat — mrjeeves · 2026-09-08
- Physicist's $100 Bounty Unclaimed for 20 Years Finally Solved by an AI Agent — IgorCarron · 2026-09-08