Our Take
Memory systems only work if they stay fresh; OpenAI's framing of 'Dreaming' suggests passive recall, but the real test is whether users can actually audit and correct what ChatGPT thinks it knows about them.
Why it matters
Repeated context-setting wastes time in workflows that span multiple sessions. If the system recalls accurately, it saves practitioners from re-explaining domain details, constraints, and stylistic preferences every conversation.
Do this week
Product leads: audit your ChatGPT usage logs this week to document which recurring preferences (output format, jargon, constraints) ChatGPT currently forgets, so you can measure whether memory cuts setup friction.
ChatGPT gains memory across conversations
OpenAI announced a memory system for ChatGPT that retains user preferences and context between separate conversations. The feature lets ChatGPT remember settings, working styles, and domain-specific details without users restating them each session. The company frames this as 'Dreaming', suggesting the model consolidates patterns across past interactions (company-reported).
The memory system operates at the user level, not the conversation level. Once enabled, ChatGPT should recall facts about your role, communication style, and functional priorities across new chats. OpenAI has not published technical details on retention scope, user controls for deletion, or how the system handles conflicting preferences across conversations.
Context setup is friction; memory reduces it
Most multi-turn workflows begin with preamble: explaining your role, domain constraints, output format preferences, and decision criteria. That overhead compounds across dozens of chats per week. If memory works, practitioners skip the preamble and jump to work.
The catch: memory is only useful if it is auditable. Users need to see what ChatGPT thinks it knows, correct errors, and delete stale preferences. Opaque recall erodes trust. OpenAI has not detailed user controls or transparency mechanisms yet, which means early adopters are running an implicit experiment on memory accuracy and privacy.
Treat memory as opt-in training data
Enable memory only for accounts and contexts where you control the data you expose. ChatGPT's memory system will learn from every interaction; if you use it for sensitive client work, research drafts, or competitive analysis, that training signal persists in your account model. Audit what you share in early chats before memory activates. Log which preferences ChatGPT infers (output style, domain terminology, risk appetite) and flag mismatches manually until you trust the system. Do not assume memory respects hard boundaries between work contexts; it is a convenience feature, not a security boundary.