“I want ChatGPT with infinite local memory” - Here’s how

Key Concepts

• MEM agent: A 4 billion parameter LLM fine-tuned for memory management, context retrieval, and markdown organization.
• MCP (Memory Connector Protocol): A server that wraps MEM agent and exposes its capabilities to other applications.
• Local Memory: Storing data locally for privacy and security.
• Context Engineering: Structuring and organizing information for efficient AI retrieval.
• Entities: Representing different areas or people in your life as separate markdown files.
• Filters: Preventing specific information from being revealed to AI models.
• Wiki Links: Connecting different markdown files to create a knowledge base.
• LM Studio: An application that allows you to run AI models locally.
• Hugging Face: A platform for sharing and discovering AI models and datasets.
• UV: A package manager that's a lot faster than pip.

1. Introduction to MEM Agent

• MEM agent allows CH GPT and CLO to have infinite memory that is 100% local and private.
• It takes context engineering to the next level.
• The project is relatively unknown.
• MEM agent provides infinite local memory for AI agents and connects memory between CHGBT, cloud, Google Docs, LM Studio, and other apps.
• This allows for reusing the same context across different tools.

2. How MEM Agent Works

• Model: MEM agent is a 4 billion parameter LLM fine-tuned for memory management, retrieval, storing of context, and organizing markdown files.
• MCP (Server): The MCP wraps the MEM agent and exposes its capabilities to other applications like cloud code, CHBD, Google Docs, or GitHub.
• Data is stored locally, ensuring privacy and preventing issues like Claude reporting inappropriate prompts.
• The 4B model can be run on most computers, unlike larger models requiring significant RAM.
• MEM agent performs well on knowledge benchmarks, rivaling models much larger in size.

3. Setting Up MEM Agent

• Cloning the GitHub Repository:
  • Copy the GitHub link.
  • Use a code editor (Cursor, VS Code, etc.) and open the terminal.
  • Type git clone [link].
  • Navigate to the new folder using cd mem-agent-mcp.
  • Choose a directory without spaces in the folder names to avoid issues.
• Using the Makefile:
  • The Makefile simplifies installation with bundled commands.
  • Run make check-uv to check if UV is installed.
  • Run make install to install LM Studio CLI.
• Running the Agent:
  • Run make run-agent.
  • Choose a model size (4-bit, 8-bit, or 16-bit) based on computer power. 4-bit is the fastest.
  • The setup will install the model from Hugging Face if it's not already installed.
  • The server will run on port 8000. Keep the terminal open.
  • If you encounter an "LMS command not found" error, run the command to add the LM Studio CLI to your path.
• Choosing a Local Memory Folder:
  • Create a new folder (e.g., "local memory") to store markdown files.
  • Open a new terminal, navigate to the MEM agent MCP folder, and run make setup.
  • Select the local memory folder as the main memory directory.
• Matching the Model Name:
  • Check the model identifier in the terminal where the agent is running (e.g., me-agent-mlx2).
  • Rename the MLX_MODEL_NAME in the configuration to match the identifier.
• Checking for Multiple Instances:
  • Run lmssp to check for multiple instances of the model.
  • Unload unnecessary instances using lms unload [model_name].
• Generating the MCP JSON:
  • Run make generate-mc-json to create the MCP.json file.
  • This file is needed to integrate MEM agent with other tools.

4. Integrating with CLO Desktop

• Download and install CLO Desktop from clo.ai/download.
• Log in with your Anthropic account.
• Open the settings in CLO Desktop and find the "Developer" section.
• Click "Edit Config" to open the cloth-desktop-config.json file.
• Update the cloth-desktop-config.json file by adding the new MCP at the bottom.
• Use cursor or any AI coding editor to help adjust this file to properly add this in the correct format.
• Use UV as the command and structure the arguments in a different way.
• Restart CLO Desktop.
• Start the MCP server by navigating to the MEM agent MCP folder in the terminal and running uv run python mcp_server/server.py.

5. Using MEM Agent with CLO Desktop

• Ask CLO Desktop to use MEM agent to save information.
• MEM agent will store the information in markdown files in the local memory folder.
• The information is stored in a context engineering way.
• Create new entities to store information about different areas or people.
• MEM agent will create new markdown files for each entity and update the main user file with wiki links.
• You can edit the markdown files directly to add or modify information.

6. Protecting Information with Filters

• Use filters to prevent specific information from being revealed to AI models.
• Run make add-filters to add filters.
• Type in the information you never want to be revealed and then type quit() to end the process.
• The filters are stored in the filters file.
• MEM agent will not reveal any information that is protected by the filters.

7. Connecting to Other Software

• MEM agent can be used with other software such as CH GPD, GitHub, and Google Docs.
• The memory_connectors subfolder in the GitHub repository contains setups for different tools.
• Use cursor or cloth code to help set up the connectors.

8. Conclusion

• MEM agent is a taste of the future of context engineering.
• It allows you to securely store your files locally and use them with cutting-edge AI models.
• It protects your private information from ever accessing the internet.
• The project is new and may have unexpected errors, but it has the potential to be very useful.