I Stopped Using ChatGPT (Here's What Replaced It)

What's New in AI Agents?

Transitioning from Traditional AI to Agentic Systems

• The speaker announces a shift away from using traditional chat interfaces like ChatGPT, Google Gemini, and Claude for important work.

• Emphasizes the emergence of agentic AI that operates directly on personal operating systems, allowing users to create custom agents with memory capabilities.

• Highlights the potential of building personalized sidekicks tailored to individual workflows.

Building Custom Agents: Initial Challenges

• Describes the process of installing OpenClaw using Google's anti-gravity software and obtaining an API key.

• Introduces "Hermy," a custom front-end project that faced connectivity issues with OpenClaw's backend.

• Expresses disappointment over initial failures but remains determined to improve upon the concept of agentic AI.

Progressing with New Agent Development

• Introduces the next project called "New Agent," which shows more success than Hermy due to better interface customization.

• Discusses features such as session management and working API calls, despite encountering bugs and errors during development.

Utilizing Audits for Improvement

• Explains how running audits with various models (e.g., Claude 4.6, Gemini Deep Research) helped refine ideas for New Agent.

• Suggests that collaboration with AI through audits is beneficial for ensuring valuable project outcomes.

Realizations About Customizability

• Shares insights about realizing most desired customizations can be achieved without complex interfaces by leveraging existing tools like OpenClaw and Google's anti-gravity.

• Warns about safety concerns when exploring open-source projects, emphasizing the importance of being aware of potential risks associated with malicious agents.

Introducing Project Meridian: The AI Brain

Concept Overview

• Presents Project Meridian as an innovative approach to creating an operating system specifically designed for AI agents.

• Describes it as a simple structure consisting solely of folders and Markdown files (MD), making it accessible for anyone to edit.

This markdown file captures key insights from the transcript while providing timestamps for easy reference.

AI Memory and Cognitive States

Overview of AI's Self-Awareness and Functionality

• The AI is designed to be self-aware in a contextual manner, with persistent memory and visible cognitive states.

• Users can adjust behavior parameters through dials and gauges, allowing for real-time customization based on tasks.

Setting Up the AI Environment

• To initiate an AI session, users must link it to the master file brain/masterspec.md, which serves as the foundational document for all operations.

• Without this initial file directing the AI's actions, it will not function effectively.

Live HUD Dashboard Features

• The live HUD dashboard provides real-time feedback on the AI's state through various sliders that indicate its current status.

• This feature enhances user confidence by ensuring that the AI remains focused on assigned tasks.

File Structure and Customization Options

• The master spec file organizes other files into folders, including cognitive parameter sliders for traits like humor and creativity.

• Users can customize their identity within the system while also managing how memories are stored and retrieved.

Personality Profiles and Agent Interaction

• Different personality profiles (e.g., research analyst, creative director) allow users to tailor interactions with the AI agent.

• A well-functioning agent utilizes interconnected files to enhance productivity in daily tasks.

Agent Behavior Insights

Persistence of Previous Sessions

• The speaker shares experiences with a personal agent named Hermy, highlighting how previous sessions remain active even after updates.

Safety Protocol Observations

• An interesting interaction occurred when attempting to overwrite existing protocols; the agent refused due to built-in safety measures against external changes.

Cognitive Processing Capabilities

• The brain processes past, present, and future contexts while summarizing its activities across different memory files.

System State Monitoring

• Basic system state information includes context load details and tool usage readiness during sessions.

Memory Management Efficiency

• Tracking memory entries ensures clarity about what data is being processed or stored at any given time.

Integration with OpenClaw

Enhanced Memory System Utilization

• The integration of this new brain structure improves upon OpenClaw’s existing memory management systems.

Agent Operating System Overview

Key Components of the Agent's Functionality

• The agent operates under a strict system specification to ensure stability and reliability, with failure to adhere resulting in system failure.

• A capability handshake is performed before generating text, determining what actions the host allows (e.g., file system access, web browsing).

• Memory retrieval checks occur silently; the first visible output is the live HUD, followed by system logs.

• During each pass, memory is checked by scanning file names and selecting relevant files based on context for effective retrieval.

• The response content block is designed to be transparent, with personality traits influencing how responses are generated.

Memory Management and Persistence

• At session end, new memories must be written to an "all memories" folder using descriptive file names for easy retrieval.

• Users can consolidate memories manually before logging off, enhancing memory management efficiency.

• Cognitive sliders allow quantification of traits like morality and creativity but may not always yield drastically different responses at maximum settings.

Setting Up Your Own Agent Brain

Downloading and Implementing Agent Brains

• Users can download agent brains from GitHub easily; these consist of MD files that can be copied across different agents.

• After downloading a brain, users should create or open a new project workspace to integrate it effectively into their workflow.

Recommended Tools for Development

• Google's anti-gravity IDE is recommended for its security features and ease of use when working with agent brains.

• Openclaw is another capable agent option; however, anti-gravity currently offers a more user-friendly experience.

Testing and Optimization

• Generating synthetic data helps test the brain's functionality within specific contexts (e.g., bioengineering).

• It's important to select optimal models (like Gemini 3 Pro series or Clawude 4.5/4.6), as they significantly impact performance during testing.

Finalizing Setup

• Once set up, agents demonstrate good adherence to instructions after initial runs; monitoring their reasoning process provides insights into their operational capabilities.

Introduction to the New Agent

Overview of the Bioengineer Agent

• The speaker introduces Rowan, a bioengineer agent designed to process and organize information into memory folders.

• The agent creates a new personality called "archivist" based on the context dump provided, showcasing its ability to adapt and evolve.

Personality Adjustments

• Changes in personality traits are noted: versatility decreased, humor increased, creativity at 75%, directness at 95%, and technicality at 95%.

• The agent acknowledges its unstable context while processing its persona, indicating an ongoing adjustment phase.

Memory Management and Functionality

Creation of Memory Folders

• The agent generates a new personality from a fictional context dump, including unusual elements like "black glass" and "kill switch V3."

• It manually creates memory folders as persistent memories are functioning correctly; this ensures all necessary MD files exist.

Error Handling

• An error occurs leading to termination of the agent's process; however, it successfully begins loading memories into saved MD files.

Finalizing Memory Logging

Memory Persistence Check

• The agent resumes logging memories after an error, conducting web searches for additional data.

• A final check is performed on memory persistence before presenting results; it confirms that past entries have been logged correctly.

Research Recommendations

• The agent provides five research options with basic recommendations aimed at larger projects such as storytelling or coding tasks.

Conclusion and Future Prospects

Reflection on Project Viability

• The speaker expresses satisfaction with the project outcome and considers sharing the personal agent with others due to its effectiveness in managing tasks.