Users are building manual Markdown files to manage AI memory

The limits of built-in memory

When you start using a new AI tool, the system typically begins with no knowledge of your specific preferences or past work. While many popular chatbots now offer built-in memory features, these systems can sometimes struggle to keep different contexts separate. A user might find that a chatbot mixes up details from a finished project with a new one, or it might cling to an old preference that is no longer relevant.

Managing these memories can be frustrating because the information is often trapped within a single platform. If you decide to switch tools, you frequently have to manually export and re-import your preferences, which rarely results in the same behavior. This lack of portability makes it difficult for people who work across multiple projects to maintain a consistent experience with their AI assistants.

To solve this, some users are moving away from native memory features in favor of an external system. This approach involves creating a portable, transparent set of files that only contain the specific information the user wants the AI to know. By taking control of the data, users can ensure that their AI tools always have the correct context for the task at hand.

Building a Markdown memory system

One way to build an external memory system is by using Markdown files. Markdown is a plain text format that is easy for humans to read and edit in any text editor, but it is also highly effective for large language models to process. This setup avoids the need for complex knowledge graphs or custom servers, making it accessible to people who are not engineers.

A basic system can be organized into three primary files to manage different types of information. The first is an 'about_me' file, which contains basic personal information, formatting preferences, and standing rules. The second is a 'projects' file, which provides a detailed overview of current work. The third is an 'archive' file, which stores completed projects so the AI can still reference past context if needed.

For those who find manual auditing tedious, the AI can actually help manage these files. Users can ask the chatbot to generate the Markdown content and then download and review it. This allows the user to audit every line to ensure accuracy before saving it as the official source of truth for the AI's instructions.

Connecting files to AI models

Integrating these files into existing tools depends on which version of the software a user is using. For example, the free version of Claude can read files from Google Drive. By placing the Markdown files in a Drive folder and enabling the connector, the AI can access the information directly. This allows the model to read the 'about_me' and project files whenever a new chat begins.

The free version of ChatGPT does not support reading directly from Google Drive, which requires a different method. In this case, users can copy and paste the relevant Markdown content into the chat window at the start of each session. To make this more efficient, users can also utilize the 'Custom Instructions' feature found in the settings of both Claude and ChatGPT to tell the AI to refer to these files as the source of truth.

For users on paid plans that allow the AI to edit files directly, a 'changelog' file is a useful addition. This file tracks any changes made to the memory files during a session. By keeping a record of updates, the user can maintain a clear history of what the AI knows and how that information has evolved over time.

Why manual memory is gaining traction

The primary reason for building an external system is portability. It removes the friction of moving preferences between different AI tools, allowing users to test new models without starting from zero. This is especially useful for power users who need to switch between different models depending on the specific requirements of a task.

Transparency is another major advantage of using Markdown files. When a user manages their own files, they know exactly what information the AI can see and how it is being interpreted. This is often more reliable than asking a chatbot to 'remember' things, which can sometimes lead to unpredictable results or the inclusion of redundant data that clutters the AI's focus.

Finally, this method offers much more granular control over context. While many AI tools have features to separate different projects, keeping those projects synced across multiple platforms can be difficult. A centralized Markdown system ensures that the same project details are available regardless of which tool the user chooses to open at any given moment.

Who this approach is for

This manual memory system is not necessarily the best choice for every AI user. If a person only uses a single AI tool for basic, everyday tasks, the effort required to maintain a Markdown system might be overkill. For simple queries and one-off tasks, the built-in memory of a single platform is usually sufficient and much faster to manage.

However, for people who work on multiple complex projects, this system provides significant benefits. It helps prevent the AI from mixing up details between different workstreams and ensures that specific formatting and style preferences are always followed. It turns the AI's memory into a tool that the human controls, rather than a black box that the human has to manage.

As AI tools continue to evolve, the ability to manage your own context will likely become more important. Users who want to maintain a high level of consistency across different platforms will find that these plain text files offer a reliable way to stay organized. It balances the need for AI assistance with the need for human oversight and data portability.