Trying to decide how much to trust your AI’s memory? Here’s the fast way to figure out where you actually need to be. I kept hitting the same wall: ChatGPT remembers my writing style but forgets the project decision I made yesterday. Then I came across a sharp breakdown from creator Jeff Su, who lays out the three levels of AI memory and, more importantly, explains why the auto-generated stuff is thin on purpose. I was genuinely relieved when I understood it, because it turns out my chatbot wasn’t broken. It was doing exactly what it was designed to do.
The original poster frames the whole thing as a ladder. You don’t have to climb to the top. You just have to know which rung fits your work. Let me walk you through his criteria so you can pick.
The criteria that matters
Before comparing, the expert sets up two questions that decide everything:
- Who controls what gets saved? The AI, or you?
- How wide is the boundary? Does the memory apply to your whole life, one project, or exactly the task at hand?
Every level is really just a different answer to those two questions. Keep them in mind as we go.
Level 1: Global memory
These are the account-level facts your chatbot auto-writes once you turn memory on. The catch, as the author points out, is that these notes get dragged into every single chat.
That’s why it stays shallow by design. Think about how many hats you wear: manager, parent, spouse, weekend golfer. Context that helps in one role is noise in another. So the AI deliberately keeps global memory thin, because if something wrong lands there, it poisons everything downstream.
- 👍 Great for high-level details like your role and tone that should follow you everywhere.
- 👎 Never holds enough specific context for real project work.
The creator tested this himself: after hours prepping a presentation, a fresh chat the next day had no clue what he was talking about. The workarounds (manually saying “update your memory” or connecting Google Drive) help, but you have to remember to trigger them every time. Claude and Gemini behave identically here.
Level 2: Project memory
Now the boundary tightens. The projects feature draws a fence around one workstream, so the AI can write much more specific memory without leaking it into unrelated chats. Global memory still cascades in, so you get both layers.
The person who shared this shows the upside clearly: ask his presentation project “what’s next?” and Claude correctly says version one slides are done, visual polish is next. Nice.
- 👍 Much richer context about one specific project.
- 👎 The AI is still the author. It decides what’s worth keeping.
And that’s the crack. When he asked for the full list of confirmed attendees, Claude gave a partial answer, even though he’d pasted the calendar invite a week earlier. The AI read it fine at the time, then quietly decided those names didn’t matter. Imagine it drafting a note to your manager with an outdated presentation date. Not good.
Level 3: Your own memory system
This is the level the expert clearly favors, and I get why. Here, memory stops being a black box and becomes actual plain-text files you own, read, and edit. Your AI reads them at the start of every session and updates them as you work.
All three big labs now ship tools for this: Anthropic’s Claude Cowork and Claude Code, OpenAI’s ChatGPT Work and Codex, and Gemini Spark. The mechanism the author describes is elegant:
- One small root file acts as a routing table for all your active projects.
- Say “continue the micro iPhone presentation” and the system routes to the right folder, even if you have 20+ projects.
- It loads only that project’s files, so you pick up exactly where you left off.
- When you wrap up, it separates changes it can make itself from ones needing your approval.
In his demo, Cowork read the root memory file, found the project, updated the date change and the new attendees, and even proposed turning a one-off “don’t use acronyms” correction into a permanent rule. You stay in control of what and where; the AI does the grunt work of upkeep.
The recommendation
Here’s how I’d read the creator’s advice:
- Just want your tone and role remembered? Level 1 is fine. Don’t overthink it.
- Running one recurring project? Level 2 gives you real value with zero setup, as long as you spot-check the facts.
- Doing serious, ongoing knowledge work across many projects? Level 3 is worth the effort. You get total control plus minimal manual work.
The honest tradeoff the author names: Level 3 takes real setup and works differently from the chat window you’re used to.
How to get started
- Open your current memory page (in ChatGPT: Personalization > Memory > Manage) and see how thin it really is.
- Move one active project into a Project to feel the tighter boundary.
- Audit what it remembers. Correct one wrong fact manually so you learn the limits.
- If you outgrow that, try a system tool like Claude Cowork and mount a workspace folder.
- Create one root routing file plus a folder per project, in plain text you can edit in something like Obsidian.
- End each session with “let’s wrap up” so the AI proposes rules and saves progress.
One more nugget: the mind behind this also uses Granola, an AI notepad, to capture in-person chats that transcript workflows usually miss.
The full video walks through the Cowork setup step by step, so if Level 3 is calling your name, go watch the whole thing.