The short answer
As of September 22, 2026. Products change; we re-check this page regularly. ChatGPT and Claude both changed their memory features in the last four months, so this page will date faster than most.
Memory in ChatGPT, Claude and Gemini comes with the plan, is on by default for most personal accounts, and captures what you say in that app with no setup. ChatGPT now ranks and consolidates memories in the background. None of the three lets memory you build in one assistant be used by another assistant or by your coding tools. If you do nearly all your AI work in one assistant, use its built-in memory. thinqOS fits when your work spans several AI tools, or when you want to inspect, correct and trace what an AI believes.
What each one does
ChatGPT has saved memories, reference to past chats, and a memory summary built in the background by a process OpenAI calls dreaming, which launched on June 4, 2026 for Plus and Pro in the US and is rolling out further. Since October 2025, Plus and Pro on the web rank memories by recency and frequency, and you can move one up or down. Since May 2026, Sources under an answer show which past chat or memory shaped it.
Claude stores memory as individual topic entries you can view, edit or delete; these replaced a daily summary on July 10, 2026. Chat search is a separate tool on paid plans. Sensitive topics such as health, religion and politics are left out by default.
Gemini has Saved info that you add yourself, and a memory of past chats for personal accounts. To stop it using something from a past chat, Google's documentation says to delete that chat.
thinqOS stores structured beliefs, each with confidence, salience and a source, for people and for agents, and other tools reach it over MCP. Claude Code and Codex connect through hooks that run whether or not the model decides to call thinqOS. Grok Build captures transcripts the same way, but it ignores hook-delivered briefing, so its recall depends on its own instructions file. thinqOS connects to ChatGPT and Claude.ai; in those chats the model has to choose to call it, and how often it does has not been measured.
| ChatGPT | Claude | Gemini | thinqOS | |
|---|---|---|---|---|
| Setup | None; on by default for consumer plans | None on personal plans | Personal accounts only, with Keep Activity on | Account, then connect each tool |
| Cost | Included in plan | Included in plan | Included | Separate product |
| See and edit memory | Edit summary, delete items, version history | View, edit, delete topics | Saved info list; past-chat influence removed by deleting the chat or correcting Gemini in chat | Mind views in the thinqOS app |
| Shows what shaped an answer | Yes, Sources under the answer | Chat search shown as tool calls | Ask Gemini | In the thinqOS app; one text block to other tools |
| Deleting a chat removes its memory | No | No | Yes; the documented method | Yes |
| Automatic ranking of memories | Yes, Plus and Pro | Not documented | Not documented | Salience decay and reinforcement |
| Contradictions | Background synthesis aims to reduce them; method not published | Edit to correct | Correct in chat | Recorded as links; retraction; confidence |
| Used by other assistants and coding tools | Not documented | Not documented | Not documented | Yes over MCP; connects to ChatGPT and Claude.ai |
| Export memory | Chat export; memories not stated | Manual copy from Settings, not a bundled export | Saved info in export not confirmed | Open formats |
| Admin controls | Workspace toggle, retention, Compliance API | Organization toggle | Consumer memory not on Workspace | Tenant roles; enterprise controls mostly not built |
Who should just use built-in memory
We would tell you to stay with built-in memory in these cases.
You do almost all your AI work in one assistant. Its memory is included, already on, and sees everything you say there.
You want no setup and are content for the vendor to decide what matters.
Your company has standardized on ChatGPT Enterprise or Claude Enterprise and needs admin controls today. Both let an admin turn memory on or off for the whole organization, and ChatGPT workspaces add retention settings and a Compliance API. thinqOS's enterprise controls are mostly not built.
You are switching assistants once. Claude and Gemini both offer a one-time memory import, and Gemini can also import chat history exported from ChatGPT or Claude.
Where built-in memory is ahead
Capture. Built-in memory sees every message in its own app. thinqOS captures fully only where a hook runs.
Seeing why, where you already are. ChatGPT's Sources appear under the answer in the app you are using. thinqOS's "Explain this response" view is in the thinqOS app; other tools receive its Mind context as one block of text.
Ranking and consolidation. ChatGPT ranks memories and builds its summary in the background, and OpenAI describes staying current over time as one of its design goals. thinqOS also decays and consolidates beliefs, and you can edit a belief's confidence yourself in the Mind views. It does not yet let you push an item up or down the ranking the way ChatGPT does.
Sensitive topics. Claude leaves sensitive topics out by default and never stores items such as government ID numbers. thinqOS's rules for sensitive content apply to what it infers, not to what it extracts.
Team administration. ChatGPT and Claude business plans have the admin controls described above.
Cost. Built-in memory comes with the plan. thinqOS is a separate product.
Where thinqOS is ahead
Each point below is in thinqOS's code and running today.
One memory across tools. The same Mind is reached over MCP from Claude Code, Codex and other tools, and it connects to ChatGPT and Claude.ai. The built-in memories stay inside their own vendor's apps. ChatGPT memory is not used by custom GPTs, Claude's app memory is separate from Claude Code's own memory, and Gemini's consumer memory is not available on work or school accounts. The Claude and Gemini imports are one-time copies, and Google's page describes a manual re-upload that overwrites what was imported before, not an ongoing sync.
Deleting a source removes what depended on it. When you delete a conversation in thinqOS, beliefs that came only from it are archived, and beliefs with other sources are recalculated. OpenAI and Anthropic both document that deleting a chat does not remove a memory created from it.
Confidence, contradiction and retraction as data. Each belief has a confidence value, contradictions are recorded as links, and a belief can be retracted with a reason. None of the three vendors documents confidence values or contradiction records.
Structure. Beliefs are structured claims about typed entities, with typed relations between them. None of the vendors documents entities or relations in its memory.
Open export. A Mind exports as JSON-LD, YAML-LD or a Markdown vault. ChatGPT's export page does not say whether saved memories are included, and self-service export is not available on Business or Enterprise workspaces.
Minds for agents as well as people, with beliefs shared to a chosen audience. OpenAI states that ChatGPT memories cannot be shared with colleagues.
thinqOS can import your ChatGPT and Claude exports. Today it reads the conversations, not Claude's memory file.
What is partial in thinqOS
Choosing what reaches the context window is still being built, goal pursuit is largely unbuilt, and learning from outcomes is partial. Enterprise controls such as single sign-on, role-based access, data residency and audit export are mostly not built.
Choose thinqOS when
Choose thinqOS when you use two or more AI tools, especially a coding agent and a chat assistant; when you want to see and fix why an AI believes something; when deleting a source should remove what came from it; or when agents need memory of their own as well as people.