OpenAI Release Notes

Library is rolling out to Enterprise, Edu, and Healthcare workspaces, giving members a dedicated place to find and reuse files they upload to or create in ChatGPT. Files saved to Library follow workspace retention policies.

Workspace owners can control whether ChatGPT automatically references Library files when responding. Turning off automatic referencing does not remove Library or prevent members from browsing, searching, opening, or attaching files themselves. For Healthcare workspaces, automatic referencing is off by default.

Library-specific Compliance API endpoints are also available for exporting and deleting Library files. Library does not introduce new external sharing or multiplayer collaboration.

Global admin console updates

Cloud Console now includes external app access controls for Sign in with ChatGPT. Admins can turn external application access on or off for the organization, require applications to be approved before members can use them, and approve or disable individual applications.

Codex updates

We’ve shipped a few Codex updates, including:

• Computer Use for Windows for Enterprise plans. See Computer Use, Managed configuration, and the configuration reference.
• Developer mode for Browser use in Chrome and the Codex in-app browser, where available. Developer mode lets Codex use controlled Chrome DevTools Protocol (CDP) access for performance profiling and deeper debugging of network traffic, console output, runtime errors, DOM state, and applied styles. The feature is turned off by default at the user level, and can be enabled from Codex app settings. Admins and owners can turn off the feature for their workspace from Codex cloud settings by accessing the Policies & Configurations pane. Learn more.

The release also adds /init in the app composer for generating an AGENTS.md scaffold, support for configuring per-app access controls for Computer Use on Windows, clearer usage-limit errors with workspace guidance and reset timing when available, and a new Unread chats section in the command menu. Learn more.

New features

• Added rate-limit reset banking for Plus and Pro users, including one free reset at launch and referral invitations for earning more during the current promotion. Eligible Business members can invite coworkers to earn shared workspace credits through a separate referral program.
• Added Developer mode for Browser use in Chrome and the Codex in-app browser. It gives Codex controlled Chrome DevTools Protocol (CDP) access for performance profiling and deeper debugging of network traffic, console output, runtime errors, and page state.
• Added the /init command to the app composer for creating project instructions with the same initialization workflow as the Codex CLI.
• Added customizable macOS Dock icons with light and dark Codex variants.
• Added Computer Use for Enterprise users outside the European Economic Area, the United Kingdom, and Switzerland.
• Added support for configuring per-app access controls for Computer Use on Windows.
• Added an Unread chats section to the command menu, with the most recently updated unread chat selected by default.

Performance improvements and bug fixes

• Made Browser use up to 2x faster through CDP and DOM snapshot optimizations that reduce browser round trips.
• Made command, browser, integration, and source activity summaries easier to understand, and improved how completed chats present files, automations, and other durable output.
• Improved plugin management by including workspace plugins, refreshing plugin state more reliably after installation or removal, and letting you upload a new version of an already-shared plugin without changing its access.
• Improved usage-limit errors with inline plan and workspace guidance, including reset timing when available.
• Added Cmd+Enter and Ctrl+Enter as shortcuts for submitting custom approval feedback.
• Fixed Browser use download handling and improved Developer mode recovery and diagnostics.
• Fixed scheduled automations so they honor the selected approval mode, and fixed manual project ordering, Browser tab dragging, MCP app sizing after right-pane transitions, and clickable ChatGPT thread mentions.
• Fixed issues affecting background agent tab restoration, commit and pull request message generation, sidebar pull request status updates, Codex Mobile QR pairing, remote-control MFA, remote SSH installation and connection, updater prompts, and overlay positioning at non-default zoom levels.
• Additional performance improvements and bug fixes.

Simplified controls in the model picker

We’re updating the model picker to make it easier to choose the balance of speed and reasoning effort that works best for your task.

The model picker now has the following simplified options:

• Instant
• Medium
• High
• Extra High [Pro plans only]
• Pro Standard [Pro plans only]
• Pro Extended [Pro plans only]

On iOS and Android, you can find the picker at the top of the conversation. On web, it appears directly in the message composer.

Users have the ability to decide whether Instant auto-switches to Medium for higher reasoning when required. You can turn this on or off in General > Settings.

What changed from the previous model picker:

Thinking Standard is now Medium, Thinking Extended is now High, and Thinking Heavy is now Extra High. Pro Standard and Pro Extended remain available under Pro. Thinking Light is no longer available.

Availability

These model picker updates are rolling out to Plus and Pro users on web, iOS, and Android globally today.

ChatGPT app experience updates

Improvement to charts, table of contents and full screen writing. Also fixes long-standing issues around editing messages with attachments, and various other bug fixes.

Web, iOS, and Android

• Interactive charts in answers: ChatGPT can now turn some answers into rich and interactive bar, line, pie, and scatter charts directly in the conversation. Ask for a chart, or ChatGPT may include one when it helps make trends and comparisons easier to understand.

Web

• Table of contents in longer conversations: Conversations longer than five responses can now include a table of contents, so you can scan sections and jump to the part you need.
• Full-screen writing blocks for longer-form work: Writing blocks now cover long-form writing use cases, including essays, PRDs, reports, blog posts, notes, and other document work. Longer writing can now open in a focused full-screen editor. You can save your document to the Library, so you can find, reuse, or edit documents later. The experience also adds a wider document layout, cleaner titles, a table of contents for long documents, download support, undo and redo fixes, improved save behavior, and clearer loading states.
• Send an email directly from within a chat: If you’ve connected Gmail or Outlook, you can now ask ChatGPT to draft and send emails directly from the same conversation. ChatGPT will create a draft and you can choose to send it without leaving ChatGPT. Sending emails is available on the web for users on Plus, Pro, Business, and Enterprise plans with Gmail or Outlook connected.

iOS

• Edit messages with attachments: You can now edit messages that include attachments, so you can revise your prompt without starting over.
• Long-press send to choose a model: Users on paid plans can now long-press the send button to choose a model for a one-off message without changing their default model.
• Autocorrection applies before sending: Last-second auto-corrections now automatically apply as before your message is sent to avoid typos.
• Immediate image previews after sending: When you upload an image on iOS, ChatGPT used to show a grey box after you sent your message. Now it shows the preview right away so that it’s clear which photo you sent.
• Cleaned up logged-out mobile web experience: Logged-out users now see a composer in the middle of the screen to avoid being blocked by popups. The sidebar has also been simplified.

Android

• Cleaner edit-message flow: Android now makes it clearer when you’re editing a previous message.

New Features

• The /app command can now hand off the current CLI thread into Codex Desktop on macOS and native Windows, and Windows workspace launches can open directly into Desktop instead of stopping at a manual prompt. (#25638, #26500)
• Local image attachments and standalone image generations now expose their saved file paths to the model, which makes follow-up edits and file references more reliable. (#25944, #25947)
• Reasoning effort selection is more flexible: the TUI adds fallback shortcuts for terminals that miss Alt bindings, and model-defined effort levels now flow through in the order advertised by the model. (#25623, #26444, #26446)
• App-server integrations can now read account token usage, and Codex auth supports v2 personal access tokens in CLI and app-server flows. (#25344, #25731)
• Plugin automation got richer structured output: add/remove and marketplace commands support --json, plugin list JSON includes marketplace source, and plugin detail data now exposes default prompts, remote MCP servers, and unavailable app templates. (#26631, #26417, #25887, #26453, #26317)

Bug Fixes

• Goal workflows are more predictable: multiline paste in /goal edit no longer submits early, idle auto-turns stay out of Plan mode, and goals stop auto-continuing after terminal turn failures. (#26047, #26147, #26690)
• Forked threads now keep user-renamed titles instead of falling back to the original first-prompt name. (#26075)
• The TUI no longer adds extra blank space while streaming, and cancelled prompts reopen with the cursor at the end so you can keep editing naturally. (#26636, #26457)
• TUI config write failures now show the underlying cause, making validation problems and read-only filesystem issues much easier to diagnose. (#26537)
• Startup is more resilient across environments, with support for /usr/bin/bash, shorter Linux proxy socket paths, and pre-refresh of expired OAuth-backed MCP credentials. (#26538, #26553, #26482)
• Workspace instruction loading is more accurate for remote and symlinked workspaces, so the right AGENTS.md files are picked up consistently. (#26205, #26465)

Documentation

• The CLI README was refreshed to remove stale guidance and better match the current documentation flow. (#26313)

Chores

• TUI startup does less repeated plugin work by reusing discovery results and loading only hook metadata on the critical path. (#26469, #26272)
• resume --last now finds the newest matching session through the state DB first, which speeds up restore on large local histories. (#26462)
• Large MCP/Ollama streams and long message histories process much faster thanks to optimized byte scanning. (#26265)

Changelog

Full Changelog:
rust-v0.137.0...rust-v0.138.0
nit: small prompt update for MAv2 (#26179) @jif-oai
feat: guard git enrichment (#26175) @jif-oai
Fix multiline paste in /goal edit (#26047) @etraut-openai
core: stop threading SandboxPolicy through exec (#25700) @bolinfest
[codex] Copy user Bazel settings into Codex worktrees (#25925) @anp-oai
[codex] Pin Python SDK to runtime 0.137.0a4 (#26216) @aibrahim-oai
Preserve remote plugin default prompts (#25887) @ericning-o
Expose local image paths to models (#25944) @won-openai
[profile-switcher][rust] -- [1/2] Add app-server account session protocol (#25469) @dhruvgupta-oai
Fix forked thread name inheritance (#26075) @etraut-openai
Restore Windows coverage for code-mode image generation exposure (#25960) @won-openai
[codex] Split Python runtime release workflow (#26226) @aibrahim-oai
feat: catalog multi-agent v2 config (#26254) @jif-oai
Rewrite oversized tool outputs during remote compaction (#26251) @pakrym-oai
codex-pr-body: avoid confidential references (#26260) @anp-oai
Use Windows setup marker as completion signal (#26074) @abhinav-oai
log plugin MCP server names (#26002) @chrisdong-oai
fix(tui): add reasoning effort fallback shortcuts (#25623) @fcoury-oai
feat(tui): add /app desktop handoff (#25638) @fcoury-oai
cli: add package path from install context (#26189) @bolinfest
[codex-analytics] report compaction request token counts (#25946) @rhan-oai
feat: catalog multi-agent v2 config (#26252) @viyatb-oai
[codex] Deduplicate skill load warnings (#26698) @xl-openai
Remove just bench-smoke from just test. (#26716) @anp-oai

App permissions for connected apps

Workspace admins can now use App permissions, where available, to choose when ChatGPT asks members before using connected apps. Admins can set a workspace-wide default and choose a different permission for individual apps, with options such as Always ask, Any changes, and Important actions.

Important actions is the default and allows ChatGPT to read from apps automatically while asking before actions that may have a meaningful effect outside ChatGPT, expose sensitive information, or be difficult to undo. App permissions replace the previous Action consent setting where available, while app access, RBAC, and Action control remain separate admin controls.

Learn more: Apps in ChatGPT; Admin Controls, Security, and Compliance in apps.

New Features

• Code mode can now call standalone web search directly, including from nested JavaScript tool calls, and receive plaintext search results. (#26719)
• Tool and connector input schemas now preserve oneOf and allOf, and large schemas keep more shallow structure when compacted, improving compatibility with richer MCP tools. (#24118, #27084)
• codex doctor now includes editor and pager environment details in the local report while redacting raw values in JSON output. (#27081)
• Plugin marketplace automation is more informative and responsive: codex plugin marketplace list --json now includes each marketplace source, and plugin lists can return from the cached remote catalog before refreshing in the background. (#27009, #26932)

Bug Fixes

• codex resume --last "..." and codex fork --last "..." now treat the trailing argument as the initial prompt instead of misreading it as a session ID. (#26818)
• MCP startup warnings from subagents now stay in the thread that owns them, avoiding duplicate parent-thread alerts and stuck startup spinners in the TUI. (#26639)
• Image edits now use the exact referenced image file paths instead of guessing from conversation history, so attached-image edits land on the intended input. (#26486)
• Bare URLs with ~ in the path are now linkified end to end in the TUI instead of being truncated before the tilde. (#27088)
• Thread resets such as /new, /clear, and /fork no longer drop cloud-managed requirements or feature flags during TUI config reloads. (#25177)
• Sandbox execution now preserves approved escalation decisions and enforces configured proxy-only networking more consistently. (#24981, #27035)

Chores

• Release builds once again publish separate symbol archives with line tables, improving post-release crash symbolication without bringing back the earlier full-debug build slowdown. (#26202)
• The embedded V8 toolchain was updated to rusty_v8 149.2.0. (#26464)

Changelog

Full Changelog: rust-v0.138.0...rust-v0.139.0

fix(remote-control): preserve enrollment on generic websocket 404s (#26741) @apanasenko-oai

fix(core-plugins): send Codex product SKU to plugin-service (#26804) @ericning-o

build(v8): update rusty_v8 to 149.2.0 (#26464) @cconger

ci: use bazel environment for BuildBuddy secret (#26895) @bolinfest

fix: preserve approval sandbox decisions in unified exec (#24981) @bolinfest

fix(tui): accept prompts with resume and fork (#26818) @fcoury-oai

deps: update starlark to 0.14.2 (#24820) @bolinfest

fix(tui): scope MCP startup status by thread (#26639) @fcoury-oai

[codex] Enable standalone web search in code mode (#26719) @rka-oai

feat: add v2 agent residency lru (#26632) @jif-oai

Ignore proc-macro-error2 advisory (#26974) @jif-oai

feat: count V2 concurrency by active execution (#26969) @jif-oai

Rename multi-agent v2 close_agent to interrupt_agent (#26994) @jif-oai

Avoid reopening v2 descendants on resume (#26997) @jif-oai

[codex] Exclude external tool output from memories (#26821) @rka-oai

[codex] Restore release symbol artifacts with line tables (#26202) @nornagon-openai

fix(app-server): avoid blocking connection cleanup (#26852) @apanasenko-oai

Add HTTP window ID to Responses client metadata (#26923) @ningyi-oai

[codex-analytics] report compaction analytics details (#26680) @rhan-oai

[codex] Speed up external agent session imports (#26637) @stefanstokic-oai

[plugins] Expose marketplace source in marketplace list JSON (#27009) @mpc-oai

ci: template custom runner names by repo (#27024) @bolinfest

fix: preserve auto review across config and delegation (#26230) @viyatb-oai

[codex] Clarify PR babysitter state mutations (#27038) @anp-oai

[codex] Calm multi-agent v2 usage prompts (#27037) @jif-oai

Pair thread environment settings (#26687) @pakrym-oai

cli: add -P sandbox permissions profile alias (#27054) @bolinfest

Enforce configured network proxy in codex sandbox (#27035) @viyatb-oai

Route image edits through referenced file paths (#26486) @won-openai

[codex-analytics] stop sending codex error subreason (#27060) @rhan-oai

[codex] Require complete main-agent skill reads (#27044) @fchen-oai

feat: support oneOf and allOf in tool input schemas (#24118) @celia-oai

[codex] Prune stale curated plugin caches (#26934) @xl-openai

Use cached remote plugin catalog for plugin list (#26932) @xl-openai

[codex] Add OTEL counter descriptions (#26091) @richardopenai

feat(doctor): report editor and pager environment (#27081) @fcoury-oai

chore: preserve one more schema layer during large tool compaction (#27084) @celia-oai

Add typed file URIs (#26840) @anp-oai

fix(tui): linkify complete bare URLs with tildes (#27088) @fcoury-oai

Show effective sandbox modes in /debug-config (#27068) @canvrno-oai

Add extra config to StoredThread, leave empty for now (#27092) @kumquatexpress

Update web search citation prompt (#27096) @yuning-oai

Preserve cloud requirements across TUI thread resets (#25177) @canvrno-oai

[codex] Remove remote compaction failure log (#27106) @pakrym-oai

New features

• Added Migrate to Codex flows for importing supported setup from Claude Code and Claude Cowork, including during onboarding.
• Revamped plugins screen with separate tabs, marketplace and category filters, keyboard navigation, and clearer install actions.
• Expanded Settings search to find options from more panels, including Git and pets.

Performance improvements and bug fixes

• Fixed goal timer overlap in narrow layouts.
• Reduced unread notifications while an active goal continues running.
• Kept review diff ordering consistent with the file tree.
• Improved window rendering on systems that don’t support translucent backdrops, including Windows 10.
• Additional performance improvements and bug fixes.

Today we’re adding five new voices that you can use with the Realtime API: Ash, Ballad, Coral, Sage, and Verse. These voices are much more expressive and you can better tune their emotions, accents, and tones to your needs. Take a listen.

We’re also making our first cost reduction for Realtime with, prompt caching:

• Text input that hits the cache costs 50% less.
• Audio input that hits the cache costs 80% less.

Basically, your typical 15 minute conversation costs about 30% less now than when we first launched the Realtime API. You can try out the new voices in the Playground right now (and be sure to use system instructions to direct their tones). Get started with the docs today—we’re looking forward to seeing and hearing what you build!

A new program to support rigorous external research on the economic impacts of AI.

AI is reshaping how people work, how businesses operate, and how ideas are created and shared. Understanding those changes will require more than anecdotes. It will require rigorous empirical research, grounded in real-world evidence and pursued by a broad community of researchers.

Today, OpenAI is launching the OpenAI Economic Research Exchange, a new platform to support high-impact external research on the economic effects of AI. Through the Exchange, selected researchers will conduct research through structured, project-based collaborations with OpenAI Economic Research, with the goal of producing credible, independent evidence on how AI is affecting workers, firms, institutions, and the broader economy.

We are seeking projects that address important questions regarding the economic impacts of AI. Proposals should explain how carefully governed, privacy-protected use of OpenAI tools can help answer these questions. The goal is to enable rigorous and credible research that would extend beyond traditional datasets alone, by utilizing OpenAI Tools and datasets. And this while maintaining clear safeguards for user privacy and responsible data use.

Selected researchers will undertake carefully scoped projects with defined milestones, data governance, and review processes. We welcome proposals from researchers with strong empirical skills and relevant expertise in applied causal inference, measurement, labor economics, productivity, firms, education, entrepreneurship, public finance, regional economics, development, inequality, or related fields.

The Exchange builds on OpenAI’s broader efforts to improve measurement of AI’s economic effects, including OpenAI Signals. By supporting a portfolio of external research collaborations, we hope to expand the evidence base available to researchers, policymakers, businesses, and the public as they navigate a period of rapid technological change.

Proposals will be evaluated based on methodological rigor, feasibility, fit with Exchange priorities, clear milestones, and the potential to contribute credible external evidence on AI's economic impacts.

For more information about the application process, see the Request for Proposals here or contact [email protected]⁠. Applications are now open here and will close July 5, 2026. We will review submissions and notify selected researchers by July 31, 2026.

Workspace agents help teams get more done together across tools. They can own entire workflows on their own, follow team processes, and be shared across your team so people can build once and use together.

We’ve also added new admin controls and visibility:

• Agent builders can set safeguards on which actions agents can take for each app enabled in their workspace.
• Business, Enterprise, and Edu admins can view workspace agent activity and usage in the admin console.

We’ve extended the free period for workspace agents until July 6, 2026. Credit-based pricing will begin on that date.

Today we’re beginning to roll out a more capable and scalable system for synthesizing memory, developed to tackle the staleness, correctness, and scalability challenges that we observe when memory is applied to the hundreds of millions of users and multi-year time horizons in ChatGPT.

Memory is what helps ChatGPT learn your preferences, projects, and constraints, allowing future conversations to start from shared context rather than from scratch.

Over the last two years, memory has grown into a critical part of the ChatGPT experience, helping ChatGPT better understand your context so it can help you accomplish meaningful goals over time. This is central to making ChatGPT more useful: knowing you, helping you, and doing more for you.

This update is available to Plus and Pro users in the US today, and will roll out to additional countries and Free and Go users over the coming weeks.

How memory has evolved

Memory first launched in April 2024 (also known as saved memories). The feature let you ask ChatGPT to remember information and carry it forward into future chats.

Saved memories were only written during the conversation and relied on strong cues to decide when to trigger memory, such as an instruction to "remember I’m traveling to Singapore in July." In practice, interacting with this system could feel like talking to someone who took a few notes, but still forgot everything that wasn’t written down. Saved memories also tend to go stale over time and eventually become incorrect or irrelevant.

In April 2025, we updated ChatGPT’s memory by giving the model the ability to reference chat context outside of the saved memories list; this was done by introducing the first version of dreaming — a method for ChatGPT to automatically curate memories in the background by referencing chat history.

In contrast to saved memories, dreaming leverages a background process that allows ChatGPT to learn from many conversations and synthesize ChatGPT’s memory state in order to always provide the freshest, most relevant context to your conversations. Dreaming also makes it easier for memory to include context that occurs naturally in conversation, without relying on explicit requests to remember something.

Over the last year, dreaming supplemented saved memories to create a step-function improvement in ChatGPT's ability to personalize responses and offset the staleness of saved memories. However, it historically was never sufficient as a standalone memory system.

Today, we are launching a significantly more capable and compute-efficient memory architecture built on top of dreaming.

The memories synthesized by dreaming are reviewable through a summary of them made visible in the memory summary page. From the memory summary, you can quickly glean the highlights of what ChatGPT knows about you, add or update information about yourself, and provide instructions on what topics ChatGPT should bring up and when. If you want to drill down into a particular area to learn more, just chat with the model.

How we evaluate memory

When we think about what "good memory" looks like in ChatGPT, a few things come to mind:

1. Carry forward useful context: You tell ChatGPT something once, and it remembers that information in your subsequent chats.
2. Follow preferences and constraints: If you describe a preference (e.g., you’re vegetarian), then ChatGPT should take actions that are consistent with that preference going forward.
3. Stay current over time: Memory should account for the passage of time. Imagine "The user is planning their birthday party for next Saturday"; eventually, Sunday arrives.

We can evaluate how ChatGPT Plus and Pro memory has improved over time with respect to each of the three memory objectives above. We do this for each of:

1. 2024: Saved memories
2. 2025: Saved memories + Dreaming V0
3. 2026: Dreaming V3

Carrying forward context

When you start a new chat with ChatGPT, you don’t have to introduce yourself from scratch. ChatGPT can save you time and build on prior context, especially for complex, long-running projects.

For example, imagine you’re using ChatGPT to shop for new camera gear that's compatible with your current camera. If you've discussed your camera setup with ChatGPT in the past, you can ask for products that are compatible with "my photography setup" and get tailored recommendations that meet your needs.

The model produces a generic response that leaves the user to do complicated compatibility checks on their own.

The model remembers the user’s camera setup and recommends a compatible product.

We can construct an eval from examples that resemble this where the model is asked to respond to a prompt that requires it to recall factual information about the user. The model is then rewarded if it responds in a way that correctly uses the relevant context. In this evaluation, the new dreaming-based system improves the model's ability to recall relevant facts.

Following preferences

Memory also helps ChatGPT respond in ways that better match your preferences and constraints.

Imagine that you’re planning a trip to Singapore. Two months before your trip, you ask ChatGPT to help with an itinerary. ChatGPT already knows from past travel planning that you enjoy wildlife photography, prefer hotels with strong AC, and enjoy a quiet dinner over a crowded bar.

The model produces a generic response that is more touristy, doesn't help with hotel booking, and largely ignores the user's interests.

The model produces a response that is personalized to the user’s interests in wildlife photography, quiet dinners, and their priorities when booking a hotel.

Preferences can take several forms:

1. Instructions for how ChatGPT should respond ("don't bring up Stan again").
2. Your personal preferences or constraints ("I’m vegetarian").
3. Implicit preferences that shape what’s relevant to you ("I live near San Francisco" → local options should be tailored to this area).

In developing the new memory system, we improved ChatGPT’s ability to apply relevant preferences from past conversations. Following the "I’m vegetarian" example above, we can evaluate whether the model correctly leverages memory to produce vegetarian-friendly dining options when a vegetarian user asks for meal prep suggestions.

Staying current over time

Time doesn’t stop when your chat ends.

Traditional memory systems can become stale. For example, you tell ChatGPT "I’m in Singapore and need a dinner recommendation for tonight." Then, time passes, your trip ends, and you wonder why ChatGPT still thinks you’re in Singapore.

With dreaming, memories are automatically updated as time passes, allowing ChatGPT to revise its memory from "You’re going to Singapore in July" to "You went to Singapore in July 2026" when the trip ends. Then, when you’re back home, ChatGPT can again provide recommendations that are tailored to your home location and time zone.

The model thinks the user is still in Singapore.

The model provides responses that are relevant to the user’s home location.

In our memory evaluations, we measure whether ChatGPT can correctly respond to prompts where the passage of time materially affects the correct answer or recommendation. Dreaming provides a substantial lift in this area:

A more scalable foundation for the future

At OpenAI, our mission is to ensure that artificial general intelligence benefits all of humanity.

While dreaming-based memory has been available to Plus and Pro users for some time, we are only now able to offer Free users a version that meets our quality bar and is practical to serve at scale. Recent improvements reduced the compute required to serve dreaming to Free users by approximately 5x, making it possible to begin rolling out dreaming to Free users over the coming weeks and to increase memory capacity for Plus and Pro users.

Looking ahead, dreaming now provides us with a shared memory foundation for all users. This update represents our most capable memory system yet, and we’ll continue improving it.

To learn more about this release and memory user controls, visit our Memory FAQ.

We’ve upgraded memory so ChatGPT can better keep your context up to date, helping responses stay more relevant. This makes memory more useful by reducing stale or contradictory saved memories and helps ChatGPT better understand your preferences, goals, and ongoing work. You can review the memories that ChatGPT may use to personalize responses through sources or in your memory summary.

Memories are now updated automatically, with ChatGPT keeping track of the details it determines are most important so it can continue building on the context you’ve already shared. If you prefer to revert to the legacy saved memories system, go to Settings > Memory > Saved memories.

This update is rolling this out to Plus and Pro users in the US today. To access it on iOS or Android, update your ChatGPT app to the latest version. We’ll expand to Free and Go plans and additional countries over the next few weeks. Users will receive an in-product notice when the update becomes available to them. For Plus and Pro users, ChatGPT can also remember more useful context, with twice as much memory capacity.

Lockdown Mode is now available to all logged-in users across account types and workspaces. It is an optional opt-in advanced security setting that limits access to the web and external services to help reduce the risk of data exfiltration from prompt injection attacks.

When Lockdown Mode is on, ChatGPT restricts network-enabled capabilities such as live web browsing, deep research, agent mode, file downloads, and some web-derived image support. Personal users can turn it on from Settings > Security. Workspace admins can configure access for members through workspace settings and role-based access controls.

Learn more: Lockdown Mode