Nathaniel Whittemore

Not even a judgment call, just surprise."

Now, heading on over to a completely different type of topic, OpenAI has shipped a huge update for their memory system, which they're calling Dreaming.

Now, some version of memory has been available in ChatGPT for a little over two years and has already come a long way.

Early versions of memory were very manual and pretty clunky, relying on a list of saved memories.

Users often needed to tell the chatbot to remember specific things and actively cull useless information from the list.

Indeed, one of the big challenges of memory is when it remembers details that are no longer relevant.

Last April, OpenAI integrated the first elements of the system that would become Dreaming.

It allowed ChatGPT to actively curate memories in the background, slowly building a more accurate picture of the user's preferences.

This upgrade made the process feel more natural and continuous, eliminating many of the early pain points.

With this release, OpenAI said that they have made the memory system much more capable and compute efficient.

Individual saved memories are gone.

Instead, the dreaming system will maintain a summary that provides richer context about the user.

The summary is fully accessible to the user and can be edited directly to make corrections or add more information.

OpenAI provided a simple example of where memory is useful in the context of asking ChatGPT about buying peripherals for a photography setup.

Without memory, ChatGPT provides generic information and makes standard recommendations.

With memory, the chatbot can tailor its suggestions to the gear the user already owns.

Now, OpenAI devised a new benchmark to test their system based on asking questions that required the model to recall relevant facts.

With the 2024 version of memory, which again was just the saved list of facts, the model succeeded on 41.5% of tasks.

The 2025 version, which added the early version of Dreaming, kicked that success up to 67.9%.

With the version of Dreaming announced this week, OpenAI found the model could succeed at 82.8% of tasks that required the recollection of relevant facts.