5 ms·

Compaction is just what Claude Code has done forever, right?

• enraged_camel 11mo ago
  I am also trying to understand the difference between compaction, and what IDEs like Cursor do when they "summarize" context over long-running conversations. Is this saying that said summarization now happens at the model level? Or are there other differences?
  • baby 11mo ago
    Codex couldnt do what claude did before when reaching full context window
  • typpilol 11mo ago
    Afaik, there's no difference besides how aggressive or not it is. But it's the same concept. Taking tokens in context and removing irreverent ones by summarizing, etc
• GardenLetter27 11mo ago
  I think the point here is not that it does compaction (which Codex also already does) - but that the model was trained with examples of the Codex compaction, so it should perform better when compaction has taken place (a common source for drops in performance for earlier models).
  • EnPissant 11mo ago
    Codex previously did only manual compaction, but yeah, maybe some extra training for compaction, too?
• baby 11mo ago
  Yes. It was missing in codex until now
• d4rkp4ttern 11mo ago
  My understanding is that they trained it to explicitly use a self-prune/self-edit tool that trims/summarizes portions of its message history (e.g. use tool results from file explorations, messages that are no longer relevant, etc) during the session, rather than "panic-compact" at the end. In any case, it would be good if it does something like this.