Containerized neural memory that learns at test time using the Titans/TTT architecture.
Unlike traditional AI memory (RAG) that stores and retrieves embeddings, this system learns during inference - its neural weights update with every interaction.
Try it now: HuggingFace Space
# Install locally
pip install -e ".[dev]"
# Run MCP server
python -m src.mcp_server- Test-Time Training: Memory updates via gradient descent during inference
- MCP Interface: Standard Model Context Protocol for tool integration
- Surprise Scoring: Quantifies how novel an input is to the learned model
Based on Google's Titans architecture and TTT layers.
Traditional Memory: store(content) -> embedding -> retrieve(query)
Neural Memory: observe(context) -> weights update -> infer(prompt)
| Tool | Description |
|---|---|
observe |
Feed context, trigger learning |
infer |
Query using learned representations |
surprise |
Measure novelty of input (0-1) |
consolidate |
Compress patterns (like sleep) |
stats |
Get memory statistics |
# Install dev dependencies
pip install -e ".[dev]"
# Run tests
pytest
# Lint and format
ruff check src/ tests/
ruff format src/ tests/
mypy src/GitHub Actions runs on every push and PR to main:
- Lint & Format - Ruff checks and formatting
- Type Check - MyPy strict mode
- Test - Pytest with coverage
- Deploy - Auto-deploy to HuggingFace Spaces on merge to
main
MIT