Cadence
A Python library for brain-inspired equilibrium networks that learn from live experience through local plasticity, memory and action.
At a Glance
Cadence Python library released under GNU GPL v3, an OSI-approved license. Alpha research software installable via pip.
Engagement
Available On
Listed Nov 2026
About Cadence
Cadence is an open-source Python library from Pragma Research for building brains that settle to an equilibrium rather than compute a forward pass. A graph of rate neurons relaxes under each observation, the action is read from that equilibrium, and synapses learn from the difference between two settles. The README describes it as alpha research software, with release v0.95.2 published on PyPI as cadence-net.
What It Is
Cadence is a library for continuing, brain-like learners built on NumPy (Python 3.11 or later). Brain.compose builds a default System 1 brain with an association region between the senses and the motor choice, a working trace carrying the last interpretation into the next settle, and an associative store of each situation's outcomes. A recall memory of recent moments is included by default. The learning rule contrasts free and nudged settles, so it uses no backward pass and no training/inference switch.
How a Brain Lives and Learns
Calling brain.live(observation, reward=...) takes the preceding action's reward and chooses the next action for one continuing stream. The brain learns while young or when aroused by surprise or want, and otherwise answers greedily from one settle. The lower-level step loop supports explicit teaching with teacher= labels, batched streams and learning from every outcome. Each action solve reports its residual, qualification and sweeps in brain.last_settlement, and an action whose settle exhausts its budget is refused without changing live state.
Memory, Imagination and Growth
Optional features include a learned code, attention-based recall, multiple recall heads and an episodic memory. brain.imagine runs private responses without changing live memory. brain.grow adds a recall memory or output slots to a trained brain, and brain.graft places trained regions of a donor brain into a larger one. Brain.save and Brain.load keep the complete continuation.
Optional System 2
Observer regions can be added through the observers argument. They read from and feed influence back to the base and motor regions and join the same settlement. The README says whether recursive feedback helps on a task is a question for evaluation.
Current Status
The README says integrated learned world prediction, repair localized to what failed and cheap stable operation are work in progress. It uses NumPy float64 on the CPU, with Numba where installed, and can use CUDA (torch) or Apple MLX GPUs for new brains in float32. It is licensed under GPL-3.0.
Community Discussions
Be the first to start a conversation about Cadence
Share your experience with Cadence, ask questions, or help others learn from your insights.
Pricing
Open Source (GPL-3.0)
Cadence Python library released under GNU GPL v3, an OSI-approved license. Alpha research software installable via pip.
- Licensed under GPL-3.0
- Python library (NumPy, Python 3.11 or later)
- Install with pip: cadence-net
- Alpha research software
Capabilities
Key Features
- Settling equilibrium brain with local learning and no backward pass
- Live continuing interaction with reward via brain.live
- Default association region, working trace, associative store and recall memory
- Optional System 2 observer regions
- Learned code, attention and multi-head recall memory
- Private imagination that does not alter live state
- Grow and graft to extend or combine trained brains
- Settlement reporting with residuals and qualification
- Save and load of full brain continuation
- Optional CUDA and MLX GPU backends
