Aether-Core is a next-generation, local AI architecture that combines the reasoning and infinite memory capabilities of Symbolic Knowledge Graphs with the fluid language generation of Sparse Neural Networks (MoE). Designed specifically for low-resource hardware, Aether-Core runs entirely on CPU or small GPUs (via 1.58-Bit Ternary Quantization) and can autonomously expand its knowledge base using a DeepSeek-Teacher API without requiring costly backpropagation for every new fact.
- Neuro-Symbolic Architecture: A lightweight Mixture-of-Experts (MoE) transformer generates human-like language, while a dynamic Symbolic Knowledge Graph serves as its source of truth and memory limitlessly avoiding hallucinations.
- Zero-Shot Knowledge Updating (API-Learning): Introduce new facts, rules, and ontology via the integrated REST API or Chat. Aether learns via "Knowledge Injection" directly into the Graph - zero model re-training and zero GPU hours needed.
- Autonomous Learning Agent: A completely background-running Observer process detects knowledge gaps during interactions or idle states, queries DeepSeek autonomously for structured facts, and populates the graph without human interaction.
- 1.58-Bit Ternary Quantization: Built-in VAE compression and {-1, 0, +1} weight optimization strictly cuts memory usage down allowing massive language models to sit effectively on consumer grade SSD/RAM/VRAM hybrids.
- Premium Web-UI with API Controller: A stunning Glassmorphism Interface out-of-the-box allows simple user interactions, built-in DeepSeek API-Key management, und background distillation triggering.
The architecture follows a strict 3-tier pipeline:
- Symbolic Memory (Graph): Stores facts, concepts, and multi-relational rules.
- Teacher Integration (DeepSeek): The background orchestrator validates data logic via a strict schema prompt and imports it directly from cloud models (Teacher -> Student pattern).
- Neural Decoder: Translates tensor-embeddings taken directly from Graph-Nodes back into highly fluent, causal linguistic sequences.
- Clone the Repo:
git clone https://github.com/YOURNAME/Aether-Core.git cd Aether-Core - Install Dependencies:
pip install -r requirements.txt
- Start the Engine:
Double click the
run_aether.batscript on Windows or execute the launcher directly:python aether_launcher.py
The Web-UI and Neural Server run concurrently and manage everything (even auto-learning loop) for you.
Aether comes with a pre-configured training distillation script. Instead of training via massive, disorganized datasets, Aether-Core queries complex conversational structures specifically tailored for your required ontology and refines its grammatical weights.
- Use the Settings Icon inside the Chat UI to enter the targeted
epochs. - Start distillation directly from the browser!
Aether-Core incorporates a state-of-the-art deterministic 3-Phase Safety Layer:
- Pre-Check Lexer: Hardcoded Entity Linker blocks Redlist tokens before they reach the graph.
- Latent Shield: Protects the memory nodes during ingestion.
- Output Scrubber: Cleanses the neural representation prior to generating responses.
- Integration of Recursive Semantic Compression (Auto-Compressing Chat histories into structural long-term facts).
- VSCode / Cursor API extension layer for seamless coding assistance.
- Fully offline Desktop-App compile utilizing PyInstaller.
MIT License - Created to push the bounds of autonomous local systems.