What's this about?
A widely-shared tip: when GPT models run inside OpenAI's Codex, they benefit from an encrypted, server-side "compaction" that summarizes old context so long tasks feel like infinite memory. Run the same GPT in a third-party harness (like Pi) and you lose that by default — explaining quality gaps people blame on the model.
What exactly is "context compaction"?
An LLM has a limited working memory (the context window). Compaction summarizes older conversation into compressed notes to make room — like tidying meeting minutes instead of keeping a full recording. OpenAI does this server-side and encrypted, so other tools can't see or replicate it — though a plugin can enable it.
What's the community saying?
940 likes original + Baoyu's Chinese relay
- Practical: harness choice silently changes model performance — benchmark the harness, not just the model.
- Strategic: server-side, encrypted features are a soft lock-in — the model is portable, the memory system isn't.