What's this about?
In Codex, Astra keeps running notes that persist across context windows instead of repeatedly compacting the conversation — and earlier windows stay searchable. Experimental now, default in the coming weeks.
What exactly is "compaction"?
When a conversation outgrows the model's context window, the harness summarizes older content to make room — lossy, like repeatedly photocopying a photocopy. Notes-plus-search replaces summarize-and-discard with write-down-and-look-up.
Why is this worth noting?
Long-running agent sessions currently degrade as compaction accumulates; this is a structural fix at the harness layer, and a preview of how labs are attacking the memory problem generally — the model curating its own external memory rather than growing its window.