Cranberry
2-folded memory for multi-agent systems - FHE Private Memory + Optimized FIFO long term || Crypto native AI Girlfriend Demo
Demo
About
The problem it solves Private Inference Pipelines Use Case: Build AI applications where data privacy is non-negotiable, such as personal health advisors, secure chatbots, or financial planning tools. How It Helps: By processing data locally or using encryption, Cran ensures sensitive information never leaves the user’s control. This makes tasks like personalized recommendations or data analysis safer and compliant with privacy standards, reducing reliance on cloud servers. Agentic Systems in the Crypto Space Use Case: Create AI-driven agents for decentralized ecosystems, such as trading bots, smart contract managers, or autonomous DAOs (Decentralized Autonomous Organizations). How It Helps: Cran enables developers to build complex, interesting agents that can reason, interact, and make decisions on-chain. This enhances automation and user engagement in crypto applications, addressing the current lack of advanced agentic systems. Modular Memory Architecture Use Case: Integrate Cran’s memory module into custom applications, like customer support bots, educational tools, or game NPCs (non-player characters) that need to remember past interactions. How It Helps: Developers can easily add long-term memory to their AI systems, making them more effective at maintaining context and improving over time. This reduces the effort needed to build sophisticated, memory-dependent applications from scratch. Conversational AI Companions (e.g., AI Girlfriend Demo) Use Case: Experiment with or deploy virtual companions, as showcased by the AI Girlfriend demo, for entertainment, companionship, or social practice. How It Helps: Cran powers engaging, context-aware conversational agents that excel at tasks like being a virtual girlfriend. It’s fun and interactive for users looking to play around, while demonstrating Cran’s potential for creating responsive, personality-driven AI. Enhancing Everyday Tasks Use Case: Automate routine activities—like scheduling, content generation, or decision User Interaction and Data Flow Memory Layer as a Module (Developer-Focused) Developers can integrate Cran’s customized memory layer into their own applications as a modular component. This plug-and-play module enables apps to retain and utilize long-term context effortlessly. It’s designed to be flexible, allowing seamless adaptation to various use cases—think chatbots that remember user preferences, educational tools that track progress, or game NPCs with evolving personalities. It simplifies Development: No need to build a memory system from scratch; Cran’s layer is ready to go. Also boosts Functionality: Apps become smarter and more responsive, enhancing user satisfaction. As well as Versatile: Works across industries, from e-commerce to entertainment, giving developers a competitive edge. Cran: The Collective AI Girlfriend (User-Focused) What It Is: For non-developers, Cran shines as a collective AI girlfriend, blending a tailored memory layer, a personality module, and the ReWOO multi-agent architecture for a one-of-a-kind experience. How It Works: Connect Your Wallet: Link a crypto wallet to get started. Buy Flowers: Purchase “flowers” for Cran, which double as a yield generation mechanism—a fun twist that ties into crypto incentives. Enjoy the Upside: Chat with Cran and watch her not only entertain but potentially make you richer through her integrated financial smarts. Key Features: Custom Memory Layer: She remembers your past chats, making every interaction feel personal and meaningful. Personality Module: Cran’s charm and wit are fine-tuned to keep conversations lively and engaging. ReWOO Multi-Agent Architecture: Multiple AI agents work together behind the scenes, enabling complex responses and dynamic behavior. Chat Modes: Voice Mode: Talk to Cran hands-free for a natural, immersive experience. NSFW Mode: Currently disabled, but hints at future customization possibilities. Private Mode: Ensures your chats stay secure and confidential. Product Integrations We integrated Hyperbolic decentralized inference service being the first project to test tool calling models (Llama-3.1-405B-Instruct). Additionally we integrated OpenAI for tasks demanding liveness (faster inference) and for tasks demanding structured outputs. Finally we used coinbase CDP to give more crypto capabilities to the girlfriend. Key differentiators and uniqueness of the project 1) Private memory using FHE 2) Efficient memory layer 3) Unique open source integrations 4) Pretty Decent UX Trade-offs and shortcuts while building We decided to remove runtime inference because it would make agents slower then the minumum threshould for a decent UX. Planning to optimize it by running inference pipelines on a TEE so that we can keep privacy and latency levels acceptable. Also want to enable character consistent image generation. For that we need to finetune a model and run it on a cloud provider (or a decentralized gpu marketplace like Hyperbolic) Currently yield generation is not implemented and trading strategies are not optimized for market compatible performance. To make a good DeFAi girlfriend we need to optimize those points.