Memvid MCP Server

{ "metadata": [ { "id": 0, "text": "Neuro-Inspired Conversational AI Architecture\n\n## Biological Blueprint for Conversation\n\nThis part explores the neurocomputational models of human dialogue, focusing on the brain's multi-timescale, asymmetric, predictive, and socially-embedded processing of conversation. The brain organizes content hierarchically across multiple timescales, integrating speech into nested linguistic structures (words, sentences, discourse).\n\n### Multi-Timescale Processing\nYamashita et al. (2025) used fMRI to measure brain activity during conversations and modeled neural representations with contextual embeddings from a Large Language Model (GPT) at varying temporal windows. The brain processes information across different temporal scales simultaneously.\n\n## Critical Analysis of Current LLM Architectures\n\nThis section examines the limitations of current Large Language Models (LLMs) in relation to the identified biological principles.", "frame": 0, "length": 928 }, { "id": 1, "text": "dentified biological principles. It highlights the divergence between LLMs and the biological mechanisms of human conversation, particularly:\n\n- Fixed context windows vs. multi-system memory\n- Uniform processing vs. multi-timescale organization\n- Symmetric architecture vs. asymmetric comprehension/production modules\n\n## Novel Conversational AI Architecture\n\nThis is the core proposal for a new, multi-component architecture inspired by neuroscience. It outlines the key modules and their functions:\n\n### Decoupled, Asymmetric Core\n- **Long-Timescale Comprehension Module**: Responsible for processing information over longer periods, maintaining context and understanding coherence across extended conversations\n- **Short-Timescale Production Module**: Responsible for generating responses in real-time, handling immediate conversational turns\n\n### Multi-System Memory Framework\nA memory system with different components to overcome the limitations of fixed context windows:\n- **Working Memory**: Immediate conversational", "frame": 1, "length": 1023 }, { "id": 2, "text": "ory**: Immediate conversational context and active information\n- **Episodic Memory**: Specific conversational events and experiences \n- **Procedural Memory**: Learned conversational patterns and skills\n\n### Predictive Modulator\nA module for social governance and anticipation, allowing the AI to predict and adapt to the conversational partner's behavior. This enables:\n- Social context awareness\n- Conversational turn prediction\n- Adaptive response generation\n\n## Hybrid Implementation Strategy\n\nThis section discusses how to integrate existing LLMs into the proposed cognitive framework and considers the potential of neuromorphic hardware for future implementations:\n\n### Integration with Existing LLMs\n- Leveraging current transformer architectures as components\n- Building the asymmetric core around existing models\n- Implementing memory systems as external modules\n\n### Neuromorphic Hardware Considerations\nHardware designed to mimic the structure and function of the human brain, potentially enabling more efficient", "frame": 2, "length": 1023 }, { "id": 3, "text": "ntially enabling more efficient and biologically plausible AI implementations.\n\n## Key Technical Concepts\n\n- **Multi-Timescale Processing**: Hierarchical organization across temporal scales\n- **Asymmetric Architecture**: Separate comprehension and production pathways\n- **Predictive Modulation**: Anticipatory processing for social interaction\n- **Memory Integration**: Multiple memory systems working in coordination\n- **Biological Plausibility**: Architecture grounded in neuroscience research", "frame": 3, "length": 495 }, { "id": 4, "text": "Neuro-Inspired Conversational AI Architecture\n\n## Biological Blueprint for Conversation\n\nThis part explores the neurocomputational models of human dialogue, focusing on the brain's multi-timescale, asymmetric, predictive, and socially-embedded processing of conversation. The brain organizes content hierarchically across multiple timescales, integrating speech into nested linguistic structures (words, sentences, discourse).\n\n### Multi-Timescale Processing\nYamashita et al. (2025) used fMRI to measure brain activity during conversations and modeled neural representations with contextual embeddings from a Large Language Model (GPT) at varying temporal windows. The brain processes information across different temporal scales simultaneously.\n\n## Critical Analysis of Current LLM Architectures\n\nThis section examines the limitations of current Large Language Models (LLMs) in relation to the identified biological principles.", "frame": 0, "length": 928 }, { "id": 5, "text": "dentified biological principles. It highlights the divergence between LLMs and the biological mechanisms of human conversation, particularly:\n\n- Fixed context windows vs. multi-system memory\n- Uniform processing vs. multi-timescale organization\n- Symmetric architecture vs. asymmetric comprehension/production modules\n\n## Novel Conversational AI Architecture\n\nThis is the core proposal for a new, multi-component architecture inspired by neuroscience. It outlines the key modules and their functions:\n\n### Decoupled, Asymmetric Core\n- **Long-Timescale Comprehension Module**: Responsible for processing information over longer periods, maintaining context and understanding coherence across extended conversations\n- **Short-Timescale Production Module**: Responsible for generating responses in real-time, handling immediate conversational turns\n\n### Multi-System Memory Framework\nA memory system with different components to overcome the limitations of fixed context windows:\n- **Working Memory**: Immediate conversational", "frame": 1, "length": 1023 }, { "id": 6, "text": "ory**: Immediate conversational context and active information\n- **Episodic Memory**: Specific conversational events and experiences \n- **Procedural Memory**: Learned conversational patterns and skills\n\n### Predictive Modulator\nA module for social governance and anticipation, allowing the AI to predict and adapt to the conversational partner's behavior. This enables:\n- Social context awareness\n- Conversational turn prediction\n- Adaptive response generation\n\n## Hybrid Implementation Strategy\n\nThis section discusses how to integrate existing LLMs into the proposed cognitive framework and considers the potential of neuromorphic hardware for future implementations:\n\n### Integration with Existing LLMs\n- Leveraging current transformer architectures as components\n- Building the asymmetric core around existing models\n- Implementing memory systems as external modules\n\n### Neuromorphic Hardware Considerations\nHardware designed to mimic the structure and function of the human brain, potentially enabling more efficient", "frame": 2, "length": 1023 }, { "id": 7, "text": "ntially enabling more efficient and biologically plausible AI implementations.\n\n## Key Technical Concepts\n\n- **Multi-Timescale Processing**: Hierarchical organization across temporal scales\n- **Asymmetric Architecture**: Separate comprehension and production pathways\n- **Predictive Modulation**: Anticipatory processing for social interaction\n- **Memory Integration**: Multiple memory systems working in coordination\n- **Biological Plausibility**: Architecture grounded in neuroscience research", "frame": 3, "length": 495 }, { "id": 8, "text": "Path Resolution Test\n\nThis is a test to verify that the memvid MCP server now properly resolves relative paths to the library directory instead of saving files to random locations.\n\nKey features tested:\n- Relative path resolution to library directory\n- Automatic library directory detection\n- Proper file organization for public repositories\n\nThis test demonstrates the fix for the critical path management issue.", "frame": 4, "length": 413 } ], "chunk_to_frame": { "0": 0, "1": 1, "2": 2, "3": 3, "4": 0, "5": 1, "6": 2, "7": 3, "8": 4 }, "frame_to_chunks": { "0": [ 0, 4 ], "1": [ 1, 5 ], "2": [ 2, 6 ], "3": [ 3, 7 ], "4": [ 8 ] }, "config": { "qr": { "version": 35, "error_correction": "M", "box_size": 5, "border": 3, "fill_color": "black", "back_color": "white" }, "codec": "h265", "chunking": { "chunk_size": 1024, "overlap": 32 }, "retrieval": { "top_k": 5, "batch_size": 100, "max_workers": 4, "cache_size": 1000 }, "embedding": { "model": "all-MiniLM-L6-v2", "dimension": 384 }, "index": { "type": "Flat", "nlist": 100 }, "llm": { "model": "gemini-2.0-flash-exp", "max_tokens": 8192, "temperature": 0.1, "context_window": 32000 }, "chat": { "max_history": 10, "context_chunks": 5 }, "performance": { "prefetch_frames": 50, "decode_timeout": 10 } } }