My AI Remembers Everything

About this video

Jeff Su argues that auto-generated AI memory is intentionally thin and breaks it into three levels. Level 1, global memory, is account-level facts and preferences that apply across every chat in tools like ChatGPT, Claude, and Gemini. Because anything saved there follows you into unrelated roles and chats, the AI keeps it high-level — for example remembering your role and writing preferences but not the details of a high-stakes presentation project — and workarounds like telling the AI to update memory or pulling from Google Drive still require you to remember to do it and cannot capture in-chat work. Level 2, project memory, scopes memory to a Projects feature for one workstream, so chats inherit both global and more specific project context, but the AI still decides what to keep and can omit details it saw, such as attendees from a calendar screenshot, or keep outdated entries. Level 3 is your own memory system built with AI systems such as Claude Cowork and Claude Code, ChatGPT Work and Codex, and Gemini Spark, where plain-text files you control act as memory. A small root memory.md routing file points to project folders, the AI reads the relevant files at the start of a session and updates them at wrap-up, proposing permanent rules for approval while auto-updating progress, giving you visibility and edit control in tools like Obsidian.

Key takeaways

• Level 1 global memory is deliberately high-level because account-wide memories apply to every chat and would pollute unrelated contexts if they were detailed.
• Level 2 project memory is more specific because it is bounded to one workstream, but the AI still authors it and can omit or retain outdated information.
• Manually telling the AI to update global or project memory works, but it depends on you noticing errors and remembering to give the correction every time.
• Level 3 uses plain-text files you own, with a root memory.md routing table that directs the AI to the correct project folder at the start of each session.
• In a Level 3 wrap-up, the AI scans the session, proposes lasting rule changes for your approval, and automatically updates progress files so the next session resumes where you left off.

Episode details: published August 12, 2026 · 14:46 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are Jeff Su’s own claims, shared for education — they are not typical or guaranteed.

← Previous: Claude Design is Insanely Easy - Become a Pro in 14 Minutes! · Next: Learn 95% of ChatGPT Work in Under 20 Minutes →

All Jeff Su videos · Jeff Su profile · Jeff Su (official site)