NuMust research

OpenAI release analysis · June 8, 2026

Dreaming makes memory alive.

A technical breakdown of ChatGPT's new memory synthesis system, evaluated by OpenAI as Dreaming V3.

Primary sources: OpenAI launch post and Memory FAQ, published June 4, 2026.

NuMust research

What actually launched

Not a new chat model. A new memory layer.

2024Saved memories

2025Saved memories + Dreaming V0

2026Dreaming V3

Naming precision

OpenAI calls the release Dreaming. “Dreaming V3” is the 2026 system label used in its evaluation comparisons.

Source: OpenAI, “Dreaming: Better memory for a more helpful ChatGPT.”

NuMust research

Published system behavior

Memory becomes a continually updated synthesis.

Context sourcesPast chats, saved memories, custom instructions, and for eligible paid users, files and connected Gmail.

→

Dreaming synthesisAutomatically updates what matters, reconciles changes, and tracks time-sensitive state across long horizons.

→

Personalized responseRelevant context is recalled when it is likely to improve the current answer, with user-visible memory sources.

Published behavior, not a disclosed model architecture. Sources: OpenAI launch post; Memory FAQ.

NuMust research

Operational interpretation

Four jobs the memory layer must perform.

01

Extract

Identify durable preferences, projects, constraints, and time-bound facts.

02

Reconcile

Resolve stale or contradictory state as newer evidence arrives.

03

Prioritize

Keep recent and frequently discussed context closer to “top of mind.”

04

Retrieve

Use personal context selectively when it improves the current response.

Inference from documented behavior. OpenAI has not published Dreaming V3's internal model, training method, or storage architecture.

NuMust research

How OpenAI evaluates memory

Three objectives define “good memory.”

Continuity

Carry forward useful facts so long-running work does not restart from zero.

Constraint following

Apply explicit and implicit preferences only when they are relevant.

Freshness

Understand that plans, locations, goals, and life state change over time.

Source: OpenAI launch post. Public charts report improvement, but complete benchmark tables are not exposed in text.

NuMust research

Product surface

More automation comes with more visibility.

Memory summary

A continuously updated high-level view of what ChatGPT remembers, with correction and history controls.

Memory sources

Responses can show which past chats, memories, instructions, files, or connected apps informed personalization.

User controls

Turn memory off, use Temporary Chat, delete sources, revert to legacy saved memories, or disconnect apps.

Rollout began June 4 for US Plus and Pro users; additional plans and countries follow over the coming weeks.

NuMust research

Why V3 matters

5×

less compute required to serve Dreaming after recent improvements, according to OpenAI.

Engineering consequence

Memory is becoming shared infrastructure, not a premium add-on.

Scale improves access, but synthesis raises harder governance questions: provenance, correction, deletion across every source, and sensitive context entering memory.

Sources: OpenAI launch post; Memory FAQ. Plus and Pro also receive twice as much memory capacity.

NuMust research

Key takeaways

The strategic shift is from stored facts to maintained user state.

Dreaming V3 makes memory fresher, more scalable, and more inspectable. The open technical question is how reliably its synthesis can remain correct, controllable, and provenance-aware over years.

Sources: https://openai.com/index/chatgpt-memory-dreaming/ · https://help.openai.com/en/articles/8590148-memory-faq