muellerberndt/cadence

A brain that learns from live experience. System 1 by default: perception, memory, plasticity and action. Optional System 2 for self-observation and slow corrections.

What it solves

Cadence provides a framework for creating "continuing brains"—AI agents that learn from experience in real-time without the traditional training/inference split. It aims to replace feed-forward deep neural networks with a system that can bootstrap a world interpretation, act in an environment, and repair failures locally as it lives through experience.

How it works

Unlike standard neural networks, Cadence uses reciprocal processing regions where neural activity "settles" into an equilibrium state. It implements a biological-inspired architecture consisting of:

  • System 1: A base brain with perception, plastic connections, and memory.
  • System 2: An optional layer of recursive feedback where observer regions monitor and influence the base brain.
  • Local Plasticity: Learning occurs through local updates based on observations and rewards, avoiding the backward differentiation (backpropagation) pass used in deep learning.
  • Memory Systems: It utilizes working traces for recent activity and associative memory for long-term retention of cue/outcome associations.
  • Imagination: A private mode that allows the brain to evaluate responses to observations without altering its live memory or state.

Who it’s for

Researchers and developers interested in biologically-inspired AI, online learning, and agents that can maintain a continuous state of existence and recover from errors through local repair.

Highlights

  • No Train/Inference Split: The brain learns and acts simultaneously in a single continuous stream.
  • Reciprocal Architecture: Uses equilibrium-based processing rather than feed-forward passes.
  • Local Repair: Repairs witnessed failures locally within the same equilibrium.
  • Biological Abstraction: Models human-like brain functions including finite capacity and arousal settings for exploration.
  • Integrated Memory: Combines working traces and persistent associative memory.

Related

  • Project
    ixaxaar/pytorch-dnc

    A PyTorch library implementing Differentiable Neural Computers (DNC), Sparse DNC (SDNC) and Sparse Access Memory (SAM) with install instructions, API usage, debugging support, and example training tasks (copy, addition, argmax).

  • Project
    meta-pytorch/captum

    Captum is a PyTorch library that provides a suite of model‑interpretability algorithms (Integrated Gradients, DeepLift, Gradient Shap, SmoothGrad, TCAV, TracIn, neuron/layer conductance, etc.) to help developers and researchers understand which inputs, neurons, or training examples drive a model’s predictions. It installs via pip/conda and works with any PyTorch model.

  • Dispatch
    Lathe: Using LLMs for Hands-On Technical Learning

    Lathe is an experimental tool that uses LLMs to generate custom, multi-part technical tutorials designed for manual implementation, helping developers learn new domains rather than skipping the work.

  • Project
    mindmuxai/brain.md

    A persistent, file-based memory layer for coding agents that stores project decisions and constraints as structured Markdown within your repository.

  • Project
    cactus-compute/needle

    Needle is a 8‑29 MB foundation model for edge devices that can do tool‑calling, structured JSON extraction, text embedding and speech‑to‑text (via the Whistle model). It uses a laddered Simple Attention architecture where every depth is a runnable sub‑model, and it ships as a single `.cact` binary with a custom C++ runtime. The Python SDK lets you decorate functions as tools, run the model locally, fine‑tune with LoRA or full‑model training, and deploy to many platforms.