My AI Remembers Everything

The Three Levels of AI Memory

AI chatbots can remember certain facts and preferences, but their built-in memory is not designed to preserve every detail of ongoing work. This video explains three levels of AI memory and how users can move from basic automated memory to a fully controlled personal memory system.

Level 1: Global Memory

Global memory stores account-level facts and preferences that can be used across different conversations. Examples include a user's role, writing preferences, and other high-level information.

The advantage is that useful personal preferences can follow the user across chats. However, global memory deliberately stays relatively thin because information saved at the account level affects every future conversation.

The video explains that detailed project information, such as presentation progress, decisions, attendees, and specific workstream context, may not be preserved reliably in global memory.

Workarounds for Global Memory

1. Explicitly ask the AI to update its memory.
2. Connect external tools such as Google Drive and provide relevant documents.
3. Use external notes and transcripts as additional context.

The video also discusses how Claude and Gemini have their own versions of global memory with similar limitations.

Level 2: Project Memory

Project memory creates a boundary around a specific workstream or recurring task. Because conversations inside the project are related, the AI can maintain more specific context than it can at the global account level.

Project memory can help an AI remember project rules, progress, and decisions. However, the AI still decides which information is important enough to remember.

This can create problems when important details are omitted or outdated information remains in memory. Users may therefore need to manually correct the AI's project memory.

The Main Limitation of Project Memory

The central limitation is that the AI remains the author of the memory. It decides what gets remembered, where it is stored, and when it is updated.

Level 3: Your Own Memory System

The third level moves AI memory from a black box into files that the user controls.

Instead of relying entirely on an AI application's internal memory, the system uses files and folders containing project information, rules, decisions, progress, and other useful context.

At the beginning of a task, the AI reads a small routing file that identifies the relevant active project. It then loads the appropriate project folder and its latest memory updates.

How Controlled AI Memory Works

1. The AI loads a small root memory or routing file.
2. The routing file identifies the relevant active project.
3. The AI opens the project's folder and reads its relevant files.
4. The AI uses those files as context for the current task.
5. At the end of a session, the AI identifies useful decisions, learnings, and progress.
6. The AI updates the appropriate memory files according to established rules.

Example: Continuing an Existing Presentation

The video demonstrates a presentation workflow where the AI reads a root memory file, identifies the relevant consulting project, and then reads the project files.

The system can recover information such as the current status of the presentation, changes to the presentation date, additional attendees, room requirements, and remaining tasks.

Because this information exists in editable files, the user can inspect and modify it directly.

Human-Controlled Memory

One of the key advantages of level three is visibility. Instead of relying on hidden AI memory, the user can open the plain-text files and edit important information directly.

For example, a presentation date can be changed manually in the project memory file, giving the user direct control over the information the AI will use in future sessions.

Automatic Session Updates

At the end of a working session, the AI can review the conversation for decisions, learnings, and progress worth preserving.

The system can propose new rules for future sessions and automatically update approved memory files with project progress.

This means the human controls what type of information belongs in memory while the AI handles much of the maintenance work.

AI Memory Across the Workspace

The routing approach can work across an entire workspace rather than only inside one project.

For example, when drafting an email about a project, the system can route the request to both the folder containing writing rules and the folder containing project-specific information. This allows the AI to combine personal working preferences with accurate project context.

Global Memory vs Project Memory vs Controlled Memory

Key Takeaways

1. Global memory is useful for broad preferences and personal information.
2. Project memory provides a tighter boundary around individual workstreams.
3. Both automated memory levels can miss important information.
4. A file-based memory system gives users direct control over what is remembered.
5. A routing file can help AI systems find the correct project context.
6. AI can handle much of the maintenance work after users define the memory rules.
7. Plain-text files provide visibility into what the AI remembers.
8. Controlled memory can help users resume projects across long gaps between sessions.

Conclusion

The video presents AI memory as a progression from simple account-level memory to project-specific memory and finally to a user-controlled memory architecture. The main idea is to combine human control over what gets remembered with AI automation for maintaining that information.