I reverse-engineered Claude Desktop's storage to give it memory (github.com)

A developer has reverse-engineered Claude Desktop’s storage system to create Mnemos, a local memory management server that enhances the chatbot's functionality by enabling memory capabilities without the need for an external API. This innovative solution intercepts and vectorizes conversation data in real-time, storing it in a local SQLite database and utilizing ONNX models for efficient indexing. The addition of a native GUI in the recently released v1.1 enables users to visually browse and search through their conversation history, presented as a semantic 3D constellation. The significance of Mnemos lies in its ability to mitigate issues associated with "context rot," where large context windows can hinder the performance of large language models (LLMs) like Claude. By implementing a selective memory approach, Mnemos allows Claude to reference only pertinent pieces of prior conversations, keeping interactions clean and reducing confusion. This improvement not only enhances the model's operational efficiency but also introduces a hybrid search mechanism that combines semantic and keyword-based results, optimizing information retrieval. The project underscores a promising direction for integrating memory structures in AI systems while maintaining user data privacy by keeping everything offline.