Aristotle 1.0
A compact experimental AI architecture for neural-symbolic reasoning, structured memory, graph reasoning, and continual learning — available as a single Python file.
Version: 1.0.3 · Framework: PyTorch · License: MIT · Status: Untrained research prototype
Quick Start
Install the runtime dependencies:
pip install torch numpy scipy safetensors
Run a smoke test:
python aristotle.py doctor --smoke
Try the included demo:
python aristotle.py demo
Use Aristotle from Python:
import aristotle
model = aristotle.create_model()
print(aristotle.capability_summary())
No package installation is required for the single-file build — place aristotle.py in your project and import it.
What Aristotle Combines
| Component | Purpose |
|---|---|
| KAN / KAT | Learn flexible nonlinear transformations |
| TTMG | Route information through internal computation states |
| EDDH | Maintain event-oriented and latent memory |
| APW | Support continual and adaptive learning mechanisms |
| GraphReasoner | Reason over structured relationships and graphs |
| Symbolic layer | Represent facts, apply rules, and verify conclusions |
| AIRD | Integrate the research model and checkpoint workflow |
The central idea is simple: use neural components to learn representations and scores, then use structured and symbolic components where explicit reasoning and verification are useful.
How It Works
A simplified flow is:
Input
↓
Neural representation
↓
Nonlinear transforms + routing + memory
↓
Graph / structured reasoning
↓
Candidate result
↓
Symbolic verification
↓
Structured output
The Math, Simply
Information is represented as vectors. A learned transformation can be written as:
h = f(Wx + b)
Instead of relying only on one fixed activation function, KAN-style components can mix several learned nonlinear functions:
f(x) = Σ pᵢ φᵢ(x)
where the weights pᵢ determine how strongly each function contributes.
Routing decisions use normalized scores such as softmax probabilities:
pᵢ = exp(sᵢ) / Σⱼ exp(sⱼ)
Graph reasoning updates a node from information arriving through its connected relations:
new_node = update(node, incoming_messages)
Memory follows the same basic idea: preserve useful previous state while integrating new information.
During training, PyTorch computes gradients of a loss L with respect to model parameters θ:
θ ← θ - η ∇θL
The symbolic layer complements this by checking explicit facts and rules. In simple terms:
Neural system: Which step looks promising?
Symbolic system: Is that step valid?
Useful Commands
python aristotle.py --help
python aristotle.py info
python aristotle.py doctor
python aristotle.py doctor --smoke
python aristotle.py demo
Checkpoints
import aristotle
model = aristotle.create_model()
aristotle.save_pretrained(model, "./checkpoint")
restored = aristotle.from_pretrained("./checkpoint")
Distribution
Use aristotle.py for the simplest experience.
Use aristotle.pyz if you need compatibility with the original internal package imports, including modules such as:
aird · apw · eddh · epistema_kan · flaneai_loss · graphreasoner · ttmg
Project Status
Aristotle 1.0.3 is an untrained research architecture, not a pretrained LLM or production-ready chatbot. Successful model construction confirms that the architecture runs; useful capabilities still depend on training, data, configuration, evaluation, and downstream integration.
The consolidated build was validated independently of the original src/ directory. The original included test suite completed with:
7 passed
Validation covered model construction, gradients, checkpoint round trips, symbolic verification, CLI startup, and bundled resources.
Files
aristotle.py Single-file consolidated source
aristotle.pyz Self-contained ZipApp
README.md Documentation
metadata.yaml Machine-readable metadata
LICENSE MIT License
Tags & Discoverability
The repository metadata includes all project tags for neural-symbolic AI, reasoning, graph systems, continual learning, memory, KAN architectures, PyTorch, research software, checkpointing, and single-file distribution.
The complete machine-readable tag set is also available in metadata.yaml. No task-specific Hugging Face pipeline_tag is declared because this release is an untrained architecture rather than a pretrained task model.
License
Released under the MIT License. See LICENSE.
Aristotle 1.0.3 — experimental AI research, packaged to be easy to try.