Himanshu Negi built "Memoir," a health application designed to manage medications, symptoms, workouts, and mood, using Cognee to implement a memory layer. The project initially relied on Gemini for chat functionality but faced significant challenges regarding data integrity and user privacy. Early iterations used hardcoded context for users, erroneously presenting a generic profile of a 25-year-old male with specific medications to every tester. Integrating Cognee allowed the AI to build a knowledge graph of user history to provide personalized insights; however, this resulted in the AI hallucinating false health data by failing to distinguish between historical testing records and current, live user information.

The resolution required an explicit system prompt directing the AI to treat the user's current profile as the sole source of truth while utilizing the memory graph only for historical context. Beyond AI-specific hurdles, the development process exposed structural gaps in managing multi-user data. Initially, the application lacked account isolation, meaning all users accessed the same shared data bucket, requiring a complete refactor of storage and API calls to implement per-account identifiers. Negi also noted that platform-specific limitations—such as free-tier token limits for AI "thinking" processes and client-side authentication constraints—created further development friction.

The project highlights that technical complexity often lies in managing system trust and infrastructure rather than the AI models themselves. For instance, the app's current build includes tradeoffs like client-side Google sign-in instead of a robust session system and the absence of server-side appointment reminders. The developer emphasizes that bridging the gap between a prototype and a product for strangers involved verifying infrastructure, such as handling email confirmation domains and managing API budget constraints. The final application remains a client-side experience focused on health data, with Negi concluding that the most critical challenges were ensuring data ownership and maintaining honest performance expectations.

Himanshu Negi built "Memoir," a health application designed to manage medications, symptoms, workouts, and mood, using Cognee to implement a memory layer. The project initially relied on Gemini for chat functionality but faced significant challenges regarding data integrity and user privacy. Early iterations used hardcoded context for users, erroneously presenting a generic profile of a 25-year-old male with specific medications to every tester. Integrating Cognee allowed the AI to build a knowledge graph of user history to provide personalized insights; however, this resulted in the AI hallucinating false health data by failing to distinguish between historical testing records and current, live user information.

The resolution required an explicit system prompt directing the AI to treat the user's current profile as the sole source of truth while utilizing the memory graph only for historical context. Beyond AI-specific hurdles, the development process exposed structural gaps in managing multi-user data. Initially, the application lacked account isolation, meaning all users accessed the same shared data bucket, requiring a complete refactor of storage and API calls to implement per-account identifiers. Negi also noted that platform-specific limitations—such as free-tier token limits for AI "thinking" processes and client-side authentication constraints—created further development friction.

The project highlights that technical complexity often lies in managing system trust and infrastructure rather than the AI models themselves. For instance, the app's current build includes tradeoffs like client-side Google sign-in instead of a robust session system and the absence of server-side appointment reminders. The developer emphasizes that bridging the gap between a prototype and a product for strangers involved verifying infrastructure, such as handling email confirmation domains and managing API budget constraints. The final application remains a client-side experience focused on health data, with Negi concluding that the most critical challenges were ensuring data ownership and maintaining honest performance expectations.