Open-Source Cadence Builds AI Agents That Never Stop Learning
Cadence is an open-source Python library for brains that learn continuously from a live stream without backpropagation, attention, or frozen weights.
- Cadence is a Python brain library that learns continuously with no backprop, attention or frozen weights.
- One
steploop handles observation, memory, prediction, action and learning with no train/inference split. - Synapses update from a local free vs nudged contrast, an equilibrium propagation variant with no backward graph.
- Planning uses imagined futures: symmetry-broken settled states compared via recorded consequences, never written to memory.
- Ships four demos (Arm, World, Connect Four, Artist) with verifiable receipts against online MLP and transformer baselines.
- Install via pip from git, NumPy-only core, optional Numba/PyTorch/MLX backends, under heavy active development.
Cadence is a new open-source Python library for building continuously learning agents, which the project calls brains. It combines learning, memory, planning, and action selection in one persistent object. Each synapse updates from the activity of its two connected neurons plus a broadcast correction signal, with no retained backward graph or global gradient exposed to neurons.
Project snapshot
| Area | Details |
|---|---|
| Runtime | Python 3.11 or later |
| Required dependency | NumPy |
| Optional backends | Numba, PyTorch, and MLX |
| Core API | GenericBrain.step |
| Learning model | Continuous online updates through equilibrium propagation |
| Project status | Under heavy development |
| Available evidence | Four end-to-end demos, event-log verifiers, and comparison runners |
One loop keeps learning
The GenericBrain.step method receives an observation and the outcome of the previous action. It then updates memory and synapses, settles the brain’s state, and selects the next action. Every call can modify the same object used for action selection, so the library has no separate pretraining, fine-tuning, inference, or frozen deployment modes.
Each brain contains recurrent cortices that settle into a joint equilibrium. An incoming reading is centered, projected through a fixed sparse expansion, and filtered by winner-take-all inhibition. Active cells record what followed previous readings, and predictions combine the records associated with the current activity. This mechanism supplies working and episodic memory inside the learning loop, without requiring a separate replay buffer.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
