Video summary

My AI Remembers Everything

Main summary

Key takeaways

Technology

Summary (technological concepts & product/method analysis)

The video argues that “auto-generated memory” features in AI chatbots often fail to remember meaningful, project-specific context, and claims this behavior is intentional rather than a bug. It then breaks down AI memory into three levels, culminating in a user-controlled “memory system” built on AI tools that read/write plain-text files.

Why auto-memory is “bad on purpose”

  • Auto-generated memory is kept intentionally thin because it affects future chats broadly.
  • If incorrect or irrelevant facts were stored globally, they could “poison” the AI’s outputs across many unrelated contexts.
  • Therefore, AI providers deliberately store less so the memory profile doesn’t degrade usefulness.

The three levels of AI memory

Level 1: Global memory (account-level)

What it is

  • Memory entries auto-generated after enabling memory.
  • Stored at the account level, so they apply across every chat.

What it typically remembers

  • High-level traits/preferences (e.g., role, writing preferences).

What it often doesn’t remember

  • Specific ongoing work (e.g., a presentation project, feedback, learnings, confirmed slide status).

Example behavior described

  • The presenter works on a presentation in one chat, but the next day a new chat has no awareness of the presentation’s context.
  • Auto-memory pages in tools like ChatGPT show little project/work-stream detail, mostly general preferences.

Why this happens

  • Because global memory cascades everywhere, the AI keeps it high-level to reduce the risk of bad context.

Workarounds

  1. Explicitly update memory in-chat (e.g., “Update your memory. I have a high-stakes presentation on October 6th.”)
    • Problem: entries still go into the thin global profile, so they can pollute unrelated chats.
  2. Connect external tools (e.g., Google Drive connectors)
    • Problem: connectors only provide retrieved material; they don’t automatically capture what happened inside prior chat sessions unless the relevant artifacts exist externally.

Applies similarly to

  • Claude and Gemini are said to behave the same way regarding global memory.

Level 2: Project memory

What it is

  • A “projects” feature that scopes memory to one work stream/recurring task.
  • Chats within a project inherit:
    • Level 1 global memory, plus
    • project-level memory generated for that work stream.

How it helps

  • Tighter boundary enables more specific reminders (e.g., presentation progress like “slides V1 are done; visual polish next”).

Main limitation

  • Still “AI authorship”: the AI decides what to remember and what to exclude.
  • Can produce confidently incomplete or outdated outputs (e.g., partially remembering confirmed attendees even when a calendar screenshot was provided earlier).

Workaround

  • Manually correct memory (e.g., updating project memory card or saying “Update project memory…” with corrected attendee info).
  • But the user must notice issues and provide corrections each time—reducing automation benefits.

Level 3: User-controlled memory system (file-based, editable)

Core idea

  • Memory becomes user-owned files (plain text), not an opaque AI-managed black box.
  • The AI system:
    • loads files at the start of each session,
    • updates them as you work, and
    • routing determines which project/task memory files to use.

Tools/labs referenced (examples)

  • Anthropic: Claude co-work / Claude code
  • OpenAI: ChatGPT work / ChatGPT Codex
  • Gemini: Gemini Spark
  • The video focuses on Claude co-work for the demonstration.

How it works under the hood (as described)

  • At the start of a task, the system loads one small routing table file (e.g., root memory.md) to identify the active project(s).
  • It then reads the relevant project folder files containing the latest memory/state.
  • Only the needed subset is loaded even with many active projects.

Demonstrated benefits

  • The AI can resume work accurately across sessions using editable files.
  • The presenter shows project state changes like:
    • slide deck readiness,
    • next steps (visual polish),
    • updating presentation date,
    • attendee list implications (booking a larger room),
    • and open items.

Critical advantage

  • No hidden memory:
    • The memory state comes from folders and plain text files that the user can open and edit (e.g., changing the presentation date in Obsidian).

Session wrap-up behavior

  • When the user ends a session (“I’m done with the session. Let’s wrap up.”):
    • the AI scans decisions/learned/progress,
    • distinguishes between items needing user approval vs. safe automatic updates,
    • then updates the relevant memory files accordingly.
  • Example: converting a one-off instruction (no acronyms in slide headlines) into a permanent rule for future sessions, and updating root memory.md so the “polish round next” status changes.

Tutorial/recommendation elements (products & onboarding)

  • Granola (sponsor): described as an AI notepad that fills note gaps by transcribing/enriching notes.
    • Works across platforms like Google Meet, Zoom, Teams.
    • Connectable to ChatGPT and Claude to fetch notes/brief and draft recaps.
    • The presenter claims it transcribes as well as Gemini and supports “AI-enhanced notes” that turn shorthand into summaries with action items.
  • Claude co-work toolkit / templates:
    • The presenter mentions a “free Co-work toolkit” with templates and a step-by-step way to build a workspace in 1 week.
  • Cowork Academy:
    • Offers a pre-built “AI command center system” to avoid building a level 3 setup from scratch.

Final comparison recap (as stated)

  • Level 1 (global): automatic, applies everywhere; too broad for real work.
  • Level 2 (project): tighter scope; still AI-controlled, so it can miss critical details.
  • Level 3 (user system): user decides what’s saved and where; AI handles maintenance (“grunt work”); best control with more setup effort.

Main speakers / sources (end)

  • Speaker: The video creator/presenter (narrator), who also demonstrates Claude co-work workflows and mentions a “Cowork toolkit/Academy.”
  • Referenced platforms/products: ChatGPT, Claude (Anthropic), Gemini, Granola, Google Drive, and note app Obsidian.

Original video