My AI Remembers Everything
Video by Jeff Su on YouTube · 15 min · uploaded August 2026
Notes and analysis by BrainTube, written from the video. Not affiliated with Jeff Su. Watch on YouTube
In short
Jeff Su sorts AI memory into three levels: account-wide global memory, project memory, and a file-based system you own, run through agent tools such as Claude Cowork. His explanation of why built-in memory is thin on purpose is the clearest part. His level three setup is sound, but it is a workflow for one person's projects, and it still depends on the AI writing good notes.
The argument
The video's main claim is that the problem with built-in memory is not capacity but authorship. In ChatGPT, Claude and Gemini the AI decides what to remember, where it lives and when it is written. Global memory stays high level because anything stored there leaks into every chat across every part of your life. Projects narrow the scope, so memory gets more specific, but the AI still chooses what to keep and you cannot edit it directly. His fix is to move memory into plain text files in folders you control: a small root file that routes each request to the right project folder, project notes the AI reads at the start of a session, and a wrap-up step at the end where the AI updates the files and asks approval for new permanent rules.
Key ideas, with timestamps
- 0:31
Global memory forgets your project
After a few hours preparing a presentation in ChatGPT, a fresh chat the next day knows nothing about it. The saved memories, under Personalization and Memory, hold only high-level facts such as his role and writing preferences. He says Claude and Gemini behave the same way.
- 1:33
Thin by design
Anything saved at account level is pulled into every future chat, so a wrong or narrow fact poisons everything. Because you wear many hats (manager, parent, golfer), context useful in one role is noise in another. That is why the AI errs toward remembering less.
- 2:35
Two level one workarounds
You can say "update your memory" with a specific fact, and the memory page updates, but you must remember to do it and the fact follows you into unrelated chats. Or you can connect Google Drive and ask it to pull a meeting transcript, which only helps if that document exists and knows nothing about earlier chat sessions.
- 5:12
Project memory
A project draws a boundary around one work stream, so the AI can write more specific memories, while global memory still cascades in. In his Claude project it correctly says version one slides are done and visual polish is next. You can view project memory but cannot edit it directly.
- 6:15
The AI is still the author
Asked for the confirmed attendees, Claude gives a partial list, even though he pasted a calendar screenshot into a chat a week earlier. It read the image at the time but chose not to keep the names. He calls this the root issue across levels one and two: the AI decides what is remembered.
- 7:48
A routing file plus project folders
At level three, memory is files. Each session starts by loading one small root memory.md that lists active projects and where they live. A request about the presentation routes to the project folder, so even with 20 or more projects only one is loaded. He names Claude Cowork and Claude Code, ChatGPT work and Codex, and Gemini Spark as tools that can run this.
- 9:21
The demo
In a fresh Cowork session, it reads the root file, finds the project in his consulting folder and reports the deck is content complete, the date moved from September 15 to October 6, and a larger room is needed. He opens the same file in Obsidian and edits the date by hand.
- 11:24
Wrap-up with approvals
After he corrects acronyms in slide headlines, he ends the session. Cowork scans for decisions and progress, proposes turning the acronym correction into a permanent rule for his approval, and updates the root file on its own so the next session starts where this one ended.
What it gets right
The distinction between capacity and authorship is the useful idea. People assume memory is weak because the model forgets. Su shows it is weak because the system is cautious about what it writes into a profile that touches every chat. That explains a lot of user frustration, and it is a better mental model than assuming memory is full or broken.
The attendee example is a good real test. The model saw the information, used it correctly at the time, and still dropped it. Anyone relying on built-in memory for work facts should expect exactly this.
The level three design has two good habits. Plain text files you can read and edit remove the black box. And the wrap-up step separates routine updates from new permanent rules that need a human yes, which is the right place to put human review.
What it leaves out
- The AI is still writing the notes. Level three moves storage into your folders, but the summaries are still drafted by the model at wrap-up. It can still drop the attendee list. The difference is that you can see and fix it, not that it stops happening.
- Raw sources. His setup stores progress notes, not the material behind them. The calendar screenshot, the meeting transcript and the research live somewhere else. When a note is wrong, you need the source to check it.
- Search. A routing file works for 20 projects you named. It does not help you find what a video said about pricing across months of material. At some point file routing needs search.
- Portability. The files are portable, but the routing behaviour lives in one tool's workspace. Moving between Claude Cowork, Codex and Gemini means setting it up in each.
- The pitch. There is a sponsor segment for Granola (about 3:39 to 5:12) and two plugs for his paid Cowork templates. The free part of the video still explains the method well.
Product names and features here are as of the video's August 2026 upload and may have changed.
How this applies to memory across AI tools
Su's three levels map well onto a practical split. Keep durable preferences in built-in memory, keep live project state in files you own, and keep the sources you learn from (videos, articles, documents) in a searchable store any assistant can reach. The last layer is the one his system does not cover, and it is what lets a note be checked against the original.
BrainTube covers that layer: it saves YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base and exposes it as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can search it. For more on why ChatGPT memory behaves the way Su describes, see does ChatGPT remember past conversations, and for carrying context between assistants, see one memory across ChatGPT, Claude and Cursor.
Who should watch it
Watch it if you use ChatGPT, Claude or Gemini for ongoing work and are frustrated that memory does not keep project details. It is aimed at knowledge workers, not developers. The explanation of why global memory stays thin (1:33 to 2:35) and the project memory failure (6:15) are the most useful parts. Skip the sponsor read from about 3:39 to 5:12. If you already keep project notes in files and use an agent tool, the level three section from 7:48 will mostly confirm what you do, though the wrap-up approval step at 11:24 is worth copying.
Frequently asked
- Why doesn't ChatGPT memory remember my projects?
- Account-level memory is added to every chat, so ChatGPT keeps it to high-level facts such as your role and preferences. Storing project details there would clutter unrelated chats. You can ask it to save a specific fact, or use a project to keep context for one work stream.
- Can I edit Claude project memory?
- In the video, Jeff Su shows that Claude project memory can be viewed but not edited directly. You correct it by telling Claude, in the project or on the memory card, to update project memory with the right information. Check current settings, as this may have changed since August 2026.
- How do I build my own AI memory system with files?
- The video's approach is one small root file listing active projects and their folders, and a folder per project holding notes the AI reads at the start of each session. At the end of a session, the AI updates the files and asks your approval before adding new permanent rules. It runs in agent tools that can read and write local files.
Keep videos like this one searchable
BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.
Go deeper
More video notes
Are you the creator and want this page changed or removed? Contact us and we will act within 48 hours.