Memory ≠ a single chat
Dreaming synthesizes memory in the background from many conversations, so deleting one chat doesn't necessarily remove what was already learned. Manage memory directly via the summary / settings.
Privacy & risks
Extended context — general, well-known considerations for a memory that persists separately from your chats. These are NOT part of OpenAI's announcement; for exact controls and policy, rely on the official Memory FAQ and your local settings.
Deletion & retention 2
The summary is highlights
The memory summary surfaces highlights of what ChatGPT knows; treat it as a curated view rather than guaranteed to list every detail.
Security 2
Personalization is persistent
When personal context is carried into every chat, be mindful about what you share — and review the summary periodically so nothing unexpected sticks around.
Prompt injection is a known category
Generally, memory + untrusted content is a discussed risk area: malicious instructions in third-party material could try to influence a model. Be cautious with content from unknown sources. (General context, not specific to this release.)
Accuracy 2
It can still be wrong
Even improved, models can misremember or confabulate. Notably, in the article the model re-verifies facts against real sources before answering — verify anything important yourself.
Context can resurface
Personalization is a double-edged sword: something mentioned earlier may shape a later answer unexpectedly. Use a Temporary Chat when you want a clean slate.
Region & policy 2
Availability varies by region
Rollout is US Plus/Pro first, with more countries and Free/Go to follow. Memory availability has historically differed by region — check what's available where you are.
Mind local data rules
Persistent personalization touches data-protection regimes (e.g. GDPR-style rights around personal data and erasure). Review your settings and regional terms before relying on it. (General context.)