Mnemo — AI Engine

Mnemo is KULVEX’s local AI inference system. It runs abliterated (uncensored) language models on your GPU via llama.cpp.

Architecture

User Message
    │
    ├── Local path: Mnemo (llama.cpp on GPU)
    │     └── Ollama-compatible API → streaming response
    │
    └── Cloud path: Claude API (Anthropic)
          └── Native tool_use → streaming response

The user toggles between local and cloud mode in the chat UI. Both paths support:

• Streaming token-by-token responses
• Tool use (17 domain agents)
• Conversation history with context building
• RAG (Retrieval-Augmented Generation) from knowledge base

Mnemo Branding

Users never see base model names. All models are branded as mnemo:

Why Abliterated?

Standard LLMs have built-in content filters that refuse certain requests. Abliterated models have these refusal mechanisms removed, giving you an AI that:

• Answers any question without moralizing
• Doesn’t lecture about safety or ethics
• Follows your instructions directly
• Acts as a tool, not a nanny

All models in the KULVEX catalog are pre-abliterated from trusted community sources (huihui-ai, mlabonne, mradermacher).

Key Components

• Model Catalog — Curated database of abliterated models with GGUF quantizations
• Model Selector — Automatic hardware-aware model selection
• llama-server — llama.cpp HTTP server (Docker container with CUDA)
• Model Router — Routes requests to the right model based on task type
• Context Builder — Constructs conversation context with system prompts and RAG