Claude vs ChatGPT Memory Decoded: Neither One Is a Record
Chad Harvey · AI for the C Suite® · September 28, 2026
This post is the written companion to the September 28, 2026 AI for the C Suite® podcast episode Capture vs. Retrieval: What AI Memory Gets Right, and Where It Loses You | AI For The C-Suite EP 83. The episode takes its own angle. This is the full record, with sources.
Between May and September 2026, both Anthropic and OpenAI rebuilt how their assistants remember you, then folded their separate products into single interfaces. The launch coverage told you what shipped. It mostly didn’t tell you what that means for the work you do every day, or how the two companies’ answers differ.
The question I keep fielding from CEOs goes something like this: if the assistant now remembers everything across every chat, project, and agent session, do I still need to keep records?
Yes. On both platforms. But the reasons differ, and the reasons matter.
This post runs both products through the same lens: what changed and when, where the memory reaches and where it stops, how each system decides what to keep, where each one loses things, and what all of that means for splitting the work between memory and documents. Claude is up first, then ChatGPT, then the two side by side.
Part 1: One Claude, One Memory
Between July 7 and September 16, Anthropic folded chat, Cowork, and memory into one Claude. Below, you’ll find what changed, where the unification stops, and a simple test for deciding what to trust to memory and what to write down.
What Changed, and When
Four dates tell the story. The memory change came first; the interface merge followed three weeks later.
| Date | What changed | Why it matters |
|---|---|---|
| Sep 16, 2026 | Chat, Cowork, and Artifacts merged into one interface. Claude routes each request itself; new slides and Docs features arrived. Pro and Max first, Free and Team later. | You no longer pick a tab. You also no longer see which mode is doing the work. |
| Sep 9, 2026 | Deadline to export legacy memory from Settings > Memory. | If migration dropped something, the recovery window has closed. |
| Aug 25, 2026 | One memory across chat and Cowork (in the cloud). Memory is now written as you talk and stored as short topic files you can read, edit, or delete. Sensitive topics are off by default. | Cowork stops needing a briefing. Memory became something you can audit. |
| Jul 7, 2026 | Cowork arrived on web and mobile, sharing one home with chat. | The groundwork for the merge. |
The mechanism change on August 25 is the one most people skimmed past. Claude used to summarize a conversation after it ended. Now it files topics continuously, while you’re talking. That’s why a correction you make once now carries everywhere.
Sources: Anthropic help center: memory · Anthropic, Aug 25 · Anthropic help center: import and export memory · TechCrunch, Aug 25 · TechCrunch, Sep 16 · Unite.AI
Where “Unified” Stops
One memory doesn’t mean one memory everywhere. Anthropic’s own documentation draws six boundaries.
Cowork memory is cloud-only. Cowork sessions that run locally on your computer don’t share it.
Projects are walled off. Each project has its own memory space, separate from other projects and from ordinary chats. That’s useful for keeping clients apart and surprising if you expected context to travel.
Some things are never stored, even if you ask: government ID numbers, financial account numbers, criminal history, and immigration status.
Sensitive topics are opt-in. Health, religion, politics, and similar categories are off by default. Turning them on works going forward only. Turn the setting off later and Claude removes the sensitive items it already saved.
Deleting a chat doesn’t delete what Claude learned from it. Memory and conversations are removed separately.
On Team and Enterprise, the owner holds the switch. Memory is off by default. Turning it off for the organization deletes every user’s memory immediately and permanently. HIPAA, public-sector, and custom-retention organizations can’t enable it at all.
Memory also stays inside Claude’s apps. Claude Code, the API, and other AI tools don’t read it.
Unified Doesn’t Mean Lossless
Most users hear “one memory” and assume nothing gets lost. Something does get lost. It just doesn’t happen where people go looking for it.
The stored files don’t decay. They’re short, durable text. The loss happens at the two ends of the process.
Capture. Claude decides, in the moment, what’s durable enough to file. It keeps the conclusion and usually drops the argument. A decision survives as one line. The three options you rejected, and why you rejected them, don’t.
Retrieval. In a new conversation, Claude sees a list of topic files with short descriptions and chooses which ones to open. If it misses the relevant file, that memory might as well not exist for that answer.
A word on sourcing. Anthropic documents what memory stores and where its boundaries are. The capture-and-retrieval framing above is my reading of how the system behaves. Anthropic doesn’t describe it that way. Treat it as a working model and test it against your own experience.
Memory for Facts, Documents for Reasoning
Memory is good at small, stable facts you could restate in one line. Documents are for anything where the path matters as much as the destination.
| Trust to memory | Write it down |
|---|---|
| Who people are, their roles, how they like things done | Why a decision was made and what was rejected |
| Standing preferences and formats | Numbers you worked through, assumptions, models |
| Company names, vocabulary, recurring constraints | Anything a board, auditor, regulator, or client could ask you to defend |
| Ongoing projects and their current status | Context that must cross project walls or leave Claude entirely |
| Corrections you want to apply everywhere | Anything a successor will need after you move on |
The test boils down to one question: would a future disagreement turn on how you got there rather than where you landed? If so, it belongs in a document.
In practice, that means closing out long working sessions with intention. When a thread produces a decision worth defending, ask Claude to write a short decision record: the decision, the options considered, the reasoning, and the date. Store it where your organization keeps records rather than leaving it in Claude alone. Then let memory carry the one-line version forward.
What to Tell a CEO on Monday About Claude
The unification is useful: less re-briefing, and one place to fix a wrong fact. It isn’t a records system. Five moves put it to work safely.
- Read your memory once, properly. Open Settings > Memory and look at the topic list. Correct or delete anything that’s wrong. The list is short, and every fix carries forward.
- Use projects as walls on purpose. Separate the clients or workstreams that must never mix. Combine the ones that should share context.
- Make a decision about sensitive topics. It’s a trade-off worth thinking through rather than a default to ignore.
- If you run a Team or Enterprise plan, talk about the off switch before anyone touches it. Turning memory off deletes everyone’s.
- Keep decision records outside Claude. Apply the test above. Anything you’d have to defend, or would hate to reconstruct, gets written down where your organization owns it.
The shift is simple. Stop treating chat history as your archive, and stop re-explaining yourself to Claude. Let memory hold the facts and documents hold the reasons.
Part 2: What ChatGPT’s Summary Leaves Out
ChatGPT’s memory changed shape in June 2026. It stopped being a list of facts you could read line by line and became a synthesized profile that ChatGPT rewrites in the background, drawing on your chats, files, and connected email. Most users noticed the new memory summary page. Fewer noticed what it implies: ChatGPT now decides what it knows about you and shows you a high-level view of the result.
The question is the same one Part 1 answered. If ChatGPT remembers you across every chat, project, and work session, what still needs to be written down? More than with Claude, because the memory is harder to inspect. And the same test applies.
What Changed, and When
OpenAI’s changes ran over five months rather than three weeks, and they arrived in a different order than Anthropic’s: visibility first, then the new memory mechanism, then the interface merge, then the boundaries. All dates come from OpenAI’s release notes.
| Date | What changed | Why it matters |
|---|---|---|
| Sep 21, 2026 | Privacy Center collects memory, personalization, and data-use information in one place (Free through Pro). | A single starting point for the audit below. |
| Aug 27, 2026 | Temporary chats can be personalized (use memory, no new memories) or not, and can be saved into history afterward. | The one clean way to keep a conversation out of memory now has an exit door back in. |
| Aug 14, 2026 | Project-only memory can be toggled on existing projects. Shared projects are always project-only. Work is unavailable in project-only projects. | Projects became walls you choose, at the cost of the agent. |
| Aug 13, 2026 | Computer History on macOS: ChatGPT can reference selected app and website activity. Off by default; Pro, Business, Enterprise (Pro only in the EEA, UK, and Switzerland). | Memory’s sources now extend beyond what you typed into ChatGPT. |
| Jul 9–16, 2026 | ChatGPT Work (the cloud agent) launched; new desktop app merged Chat, Work, and Codex with unified Recents and Projects. Cloud Work syncs across devices; local stays local. | The interface merge, the analogue to Anthropic’s Sep 16. |
| Jun 12, 2026 | “Delete and turn off memory” control on the summary page. | A one-click wipe of the profile, without deleting chats. |
| Jun 4, 2026 | “Dreaming”: memory rebuilt as a background process that synthesizes a readable, editable memory summary and revises stale items. Plus and Pro in the US first, with 2x capacity; legacy saved memories kept as an option. | The mechanism change. Memory became a profile ChatGPT writes rather than a list you keep. |
| May 5, 2026 | Memory Sources across all consumer plans: see which past chats, saved memories, custom instructions, files, and Gmail messages shaped a response. | The first audit surface, shown after the fact, one response at a time. |
Notice the relationship between June 4 and May 5. OpenAI shipped the transparency layer a month before it shipped the system that needed it. Memory Sources exist because Dreaming makes what ChatGPT knows harder to see directly.
Where ChatGPT’s Memory Stops
OpenAI’s memory help article draws the boundaries. They’re more numerous than Anthropic’s, and several are softer.
Availability varies by plan, region, and platform. Improved memory shipped to US Plus and Pro subscribers first. If a control isn’t visible, it may not be available to that account yet.
The summary isn’t the whole memory. OpenAI states that the memory summary is a high-level view that can omit details and sources. There’s no complete, itemized list under improved memory. The legacy saved-memories list is the only view that qualifies.
Chat-history memory is derived and changeable. Information drawn from past chats can change as ChatGPT updates what it finds useful. There’s no separate storage limit for it and no list to read.
Projects can be walled, but only by choice. Default memory lets project chats reference outside memory. Project-only memory blocks both directions. Shared projects are always project-only, and ChatGPT Work can’t run inside them.
Memory reaches beyond your chats. Depending on plan, sources include Library files, connected apps such as Gmail, and on macOS, Computer History. Disconnecting an app stops future access but doesn’t remove past uses.
Sensitive topics are a training preference rather than a switch. OpenAI says ChatGPT is trained not to proactively remember health details and the like unless you ask. There’s no documented list of categories that are never stored and no opt-in toggle like Anthropic’s.
Deleting a chat doesn’t delete its memory. Same as Claude. Removing something fully means deleting the memory, the chat, any Library file, and any connected-app source, in that order. Logs of deleted saved memories may be retained for up to 30 days.
Turning memory off doesn’t stop safety uses of context. OpenAI reserves limited use in rare, high-risk situations, separate from personalization.
Regulated workspaces get a different default. In ChatGPT for Healthcare and Enterprise with Regulated Workspace, improved memory is off by default, isn’t covered by the BAA, and admins choose which roles can enable it.
Personal accounts may feed training. Chats and remembered information can be used to improve models when “Improve the model for everyone” is on. Business, Enterprise, Edu, and Healthcare content isn’t used by default.
As with Claude, memory stays inside the product. The API doesn’t read it, and other assistants don’t see it.
A Summary Is a View, and a View Omits
With Claude, the misunderstanding was that “unified” meant lossless. With ChatGPT, it’s the belief that the memory summary is what ChatGPT knows. It isn’t. The summary is a synthesized view of a larger, moving body of context.
OpenAI describes three layers, and each one loses information in its own way.
The memory summary (Dreaming). A background process reads your chats, writes a profile, then revises it as things change. OpenAI’s own example is turning “you are going to Singapore in July” into “you went to Singapore in July 2026” once the trip is over. The process is built to compress. It keeps who you are and what you’re working on. It drops the argument behind each decision and, on purpose, anything stale. The summary is the most readable layer and the most lossy.
Reference chat history. ChatGPT searches past conversations when it judges that context would improve a response. OpenAI says it doesn’t do this on every message. This layer holds the most detail but has no list, no edit control, and no guarantee of being consulted.
Saved memories (legacy). The explicit list. Highest fidelity per item, but only for what was saved, and now an opt-out alternative rather than the default.
So the loss profile flips relative to Claude’s. Claude files short, durable facts and may fail to open the right one; the risk sits at retrieval. ChatGPT reads everything in the background and folds it into a profile that omits by design; the risk sits at capture. Memory Sources, shown per response, tell you what got used after the fact. They don’t tell you what got dropped.
A word on sourcing, again. The three-layer description is OpenAI’s. The claim that loss concentrates at capture is a working model built from OpenAI’s descriptions and from the mechanism’s stated purpose of reducing stale and contradictory memories. It’s reasoned rather than documented.
Memory for the Present Tense, Documents for the Past
The division of labor is the same as for Claude, with one addition. ChatGPT’s memory is built to describe you as you are now. It actively retires what has passed. That makes it a poor place to keep a record.
| Trust to memory | Write it down |
|---|---|
| Who you are, your role, current goals and preferences | Why a decision was made, what was rejected, and when |
| Ongoing projects as they stand today | The state of a project at a point in time, once it has moved on |
| Formats, tone, standing instructions (custom instructions are more reliable for these) | Numbers, assumptions, and models you worked through |
| Corrections you want applied going forward | Anything a board, auditor, regulator, or client could ask you to defend |
| Context you’re happy to have inferred from files and email | Anything that must stay out of a synthesized profile |
The test is unchanged: would a future disagreement turn on how you got there rather than where you landed? If so, it belongs in a document. ChatGPT adds a second: if you need to know what was true then, as opposed to what’s true now, memory will have revised it.
In practice, close working sessions by asking ChatGPT to write a decision record, then save it to Library or to your own system. The Library file becomes a memory source in its own right, which is the one place where ChatGPT’s design works in your favor here. Use custom instructions, rather than memory, for rules that must hold exactly.
What to Tell a CEO on Monday About ChatGPT
ChatGPT’s memory reaches further than Claude’s, into your files and email, and shows you less of what it holds. Six moves apply.
- Read your memory summary, then read your Sources for a week. The summary is the profile. Sources on individual responses are where you find out what else is being used, including files and email you may have forgotten you connected.
- Decide which layer you want. Improved memory (Dreaming) for convenience, or legacy saved memories for an itemized list you control. You can switch in Settings > Personalization > Memory. Most executives should try improved memory and know how to switch back.
- Audit what memory can reach. Connected Gmail, Library files, and Computer History are all memory sources. Disconnect anything you wouldn’t want inferred from.
- Wall your projects on purpose. Turn on project-only memory for client or confidential work and accept that Work is unavailable there. Shared projects are already walled.
- If you run a Business, Enterprise, or Healthcare workspace, read the memory section of your admin settings. Regulated workspaces default off, the BAA doesn’t cover improved memory, and role permissions decide who gets it.
- Keep decision records outside ChatGPT. Apply both tests. Anything you’d have to defend, or that must reflect a moment rather than the present, gets written down where your organization owns it.
The shift here is a matter of framing. Treat ChatGPT’s memory as a colleague’s impression of you: updated daily, never written down in full, useful and current, and nothing you’d cite in a board minute.
Part 3: Side by Side
Both companies solved the same problem: users were tired of re-briefing the assistant. Both arrived at account-level memory that reaches across chat and their agent product. From there the designs diverge, and the divergence should shape how you use each one.
| Dimension | Claude | ChatGPT |
|---|---|---|
| What memory is | A list of short topic files, written continuously as you talk | A synthesized profile (“memory summary”) rewritten in the background, plus searchable chat history |
| Can you read all of it? | Yes. Every file is listed under Settings > Memory. | No. The summary is a high-level view; chat-history memory has no list. Per-response Sources show what was used rather than what exists. |
| Who writes it | Claude files facts; you can edit any file | ChatGPT synthesizes; you can correct the summary or switch back to the legacy list |
| What happens to stale facts | They stay until edited | Revised automatically (“going to Singapore” becomes “went to Singapore”) |
| Where loss concentrates | Retrieval: the file exists but may not be opened | Capture: the profile omits and revises by design |
| Project walls | Always on. Every project is separate. | Opt-in per project. Shared projects are always walled. Walling a project disables ChatGPT Work inside it. |
| Agent mode and memory | Cloud Cowork shares memory; local Cowork doesn’t | Cloud Work shares memory; local stays local; project-only memory excludes Work |
| Sensitive topics | Off by default; explicit opt-in switch; forward-only; cleanup on opt-out | Trained not to remember proactively; no switch, no documented category list |
| Never stored | Documented list: ID numbers, account numbers, criminal history, immigration status | No documented never-stored list |
| Sources beyond chat | None documented | Library files, connected apps (Gmail), Computer History on macOS |
| Delete a chat, delete its memory? | No. Separate actions. | No. Separate actions, plus files and apps to check. |
| Org off switch | Team/Enterprise owner; turning off deletes everyone’s memory permanently | Workspace admins control availability and roles; regulated workspaces default off and sit outside the BAA |
| Training use | Not documented as a memory consideration | Personal accounts may contribute unless “Improve the model” is off; business plans excluded by default |
| Reaches other tools? | No. Neither Claude Code nor the API. | No. The API does not read it. |
The Difference That Matters Most
Claude’s memory is a filing cabinet you can open. ChatGPT’s memory is a colleague’s impression of you. Both are useful. Neither is a record.
The filing cabinet fails when the right drawer stays shut. The impression fails because it was never meant to hold detail, and it updates itself to match the present. That’s the capture-versus-retrieval flip, and it changes the advice.
With Claude, the discipline is keeping the files accurate and accepting that not every file gets read every time. Write down the reasoning and let memory carry the facts.
With ChatGPT, the discipline is checking Sources regularly, auditing what the profile can reach (email, files, desktop activity), and writing down anything that must stay fixed in time. Otherwise memory will revise it.
One Test for Both
Would a future disagreement turn on how you got there rather than where you landed? If so, it belongs in a document.
And for ChatGPT specifically: if you need to know what was true then, as opposed to what’s true now, memory will have revised it.
Everything else (the roles, the preferences, the vocabulary, the current state of ongoing work) is exactly what these systems were built for. Let them have it.
A note on method. Both reviews started with the vendor’s help center rather than the announcement, because the boundaries live in the documentation and rarely in the press release. Documented facts are stated plainly with sources. Claims about mechanism, particularly about where each system loses information, are working models built from the vendors’ own descriptions and labeled as such. Dates and plan tiers change; treat this as accurate as of September 28, 2026.
Sources: Anthropic help center, memory · Anthropic memory announcement, Aug 25 · Anthropic help center, import and export memory · Anthropic memory coverage, TechCrunch Aug 25 · Anthropic interface merge, TechCrunch Sep 16 · OpenAI Memory in ChatGPT help article · OpenAI ChatGPT release notes