Summary
Jeff Su breaks down AI memory into three distinct levels, explaining why every major AI chatbot — ChatGPT, Claude, and Gemini — seems to remember far less than users expect, and what can be done about it. The video is framed around a concrete frustration: working on a high-stakes presentation for hours, then returning the next day to find the AI has retained almost none of the project context.
Level one is global memory: account-wide facts the AI automatically saves across all chats. Su explains that this tier is deliberately kept thin because information useful in one context (work mode) can actively degrade outputs in another (personal tasks), so models err toward remembering less. Level two is project-scoped memory, available in Claude Projects and equivalent workspaces in other tools, which allows the AI to write more specific context while isolating it from unrelated chats. The limitation at both levels is the same: the AI remains the author and decides what makes the cut, which means critical details — like a confirmed attendee list pasted as a screenshot — may not survive.
Level three is the focus of the second half: user-controlled memory files that the AI reads at session start and updates as work progresses. Su points to Claude Code (Anthropic), ChatGPT Codex (OpenAI), and Gemini Spark (Google) as the current generation of tools enabling this pattern. The core principle is that memory becomes explicit files the user owns, not a black box the model maintains — shifting authorship from AI to human and making the system auditable and correctable.
📺 Source: Jeff Su · Published August 12, 2026
🏷️ Format: Tutorial Demo
