Core Concept Behind the Memory ArchitectureShort-Term Memory (STM): Managed via a rolling Redis cache buffer and database-backed sliding window session context. This keeps the immediate conversation snippet ($k$ past exchanges) active in the LLM's prompt context window for natural, fast-paced dialogue.Long-Term Memory (LTM): Handled via a Vector Database (e.g., Pinecone / MongoDB Atlas Vector Search). When a user inputs a query, relevant historical contexts, user preferences, and older abstract facts are retrieved using semantic similarity search (embeddings) and injected into the system prompt.🤖 MindSync: ChatGPT Clone with Dual-Stage Memory (MERN)MindSync is an advanced ChatGPT clone built on the MERN stack. Unlike basic conversational interfaces, this application implements a custom Dual-Stage Memory Architecture combining Short-Term Memory (STM) for localized conversation threads and Long-Term Memory (LTM) using vector embeddings to preserve abstract user facts, preferences, and long-form history across sessions.✨ Key Features🧠 Short-Term Memory (STM): Implements a sliding-window mechanism capturing the immediate context of the current session, ensuring fast, contextually accurate responses without overflowing LLM context tokens.🗄️ Long-Term Memory (LTM): Asynchronous background processing translates older conversation milestones into semantic vector embeddings. Relevant long-form memories are dynamically injected based on user intent.⚡ Real-Time Streaming: Fluid UI text-streaming using Server-Sent Events (SSE) or WebSockets for an authentic ChatGPT-style typing experience.📂 Multi-Thread Chat Management: Complete workspace interface enabling users to create, rename, and delete chat rooms, completely synchronized with persistent storage.🎨 Responsive Tailwind Design: A highly polished, responsive web layout mirroring modern AI dashboards—built with native CSS layouts and customized dark-mode themes.🏗️ Architecture & Memory TopologyThe system splits memory management into dynamic high-speed retrieval layers and long-term analytical persistence layers:Plaintext
+-------------------------------------------+ | React Client | +-------------------+-----------------------+ | (Prompt Input) v +-------------------+-----------------------+ | Express API Gateway | +-------------------+-----------------------+ | +-------------------------+-------------------------+ | (Fetch Current Window) | (Semantic Vector Query) v v +--------+----------------+ +--------+----------------+ | Short-Term Memory | | Long-Term Memory | | - Redis Buffer Cache | | - Vector Embeddings | | - Recent Context (10k) | | - Historical Facts | +--------+----------------+ +--------+----------------+ | | +-------------------------+-------------------------+ | (Injected Context Matrix) v +-------------+-------------+ | Large Language Model | +---------------------------+

The Ingestion Flow: User sends a message -> Gateway queries STM for the trailing thread window -> Gateway simultaneously sends embeddings of the message to the Vector DB to surface relevant historical contexts (LTM).Context Compilation: LTM facts + STM threads + System Prompts are compiled into an optimized prompt payload and passed to the LLM.The Consolidation Phase: Periodic background workers parse old conversations, summarize redundant threads, and write them permanently into the Vector Space while flushing expired STM entries.🛠️ Tech StackFrontend: React, Tailwind CSS v4, Vite, Axios, React Context API.Backend: Node.js, Express.js.Database & Storage: MongoDB (User accounts & message logs), Redis (Short-term cache layer).Vector Engine: Pinecone / MongoDB Atlas Vector Search (For semantic LTM retrieval).AI Integration: OpenAI API / LangChain framework (Text generation & text-embedding models).