OpenAI launched Dreaming V3 on June 4, a rebuilt memory architecture replacing the manually curated list of saved memories.
Memory in an assistant is a harder problem than it looks. Storing everything makes conversation slow and expensive, and drags in information unrelated to the current request. Storing too little makes the assistant repeat questions already answered.
Manual curation, which the new system replaces, solved that by pushing the decision to the user: you chose what the assistant should remember. It works for anyone willing to do the work, and almost nobody is.
Swapping in an automatic system relocates the problem rather than removing it. Now the system decides what to keep, and that decision becomes invisible: the user doesn't know what the assistant knows about them, or for how long.
For anyone using assistants with sensitive work data, that opacity is the part to watch. It's worth asking what is retained, where, for how long and how to delete it, before putting client information into a conversation.