Notes on Temporal ECS Neural Dynamics, October 28

October 28, 2025 (3w ago)

Computation often executes and forgets. Biology endures.

Temporal ECS Neural Dynamics model computation as a continuous, adaptive process.
Time is the medium in which structure and learning coexist.

The architecture separates responsibilities into three layers:

  1. Compute — orchestrates and executes work in parallel.
  2. Simulation — defines entities, components, and systems.
  3. Neural Dynamics — integrates experience and adapts behavior continuously.

1. Compute — parallel execution

Compute manages parallel execution of jobs, including neuron updates.

  • Parallelism: Large workloads update simultaneously.
  • Dependency resolution: Jobs execute only when prerequisites are satisfied.

Compute enables scalability, supporting large neural populations.

2. Simulation — structure and persistence

Simulation defines what exists. Entities, components, and groups provide stable, contiguous state.

  • Persistent memory: Components maintain state across cycles.
  • Pure transformations: Systems evolve state; scheduling is external.
  • Temporal continuity: State accumulates and decays naturally.

Simulation is the body of the system.

3. Neural Dynamics — continuous adaptation

Neural dynamics define how the system adapts, inhabiting the simulation and acting on its state.
Neurons and synapses exist as components within the ECS.

  • Temporal integration: Potentials evolve continuously with inputs, decay, and feedback.
  • Plasticity: Synapses adapt dynamically, encoding memory and learning.
  • Perception-action coupling: Neural outputs modify ECS state.
  • Parallel updates: Compute executes neuron updates in parallel, enabling scale.

4. Unified temporal flow

The system operates as a continuous loop:

  1. Data is streamed into the system through a network.
  2. Simulation evolves component states.
  3. Neural dynamics integrate inputs and produce outputs.

5. Computation that endures

Temporal ECS Neural Dynamics transforms computation into a living substrate:

  • Compute provides scalable, parallel execution.
  • Simulation provides persistent form and structure.
  • Neural dynamics provide adaptation and memory.

Signals propagate. Potentials integrate. Behavior emerges. Computation remembers, adapts, and persists.

Source Code: Full implementation on GitHub

References:

  • Liquid Time-constant Recurrent Neural Networks as Universal Approximators — arXiv:1811.00321
  • Liquid Time-constant Networks — arXiv:2006.04439
  • Dynamic Circular Work-Stealing Deque — PDF
  • Liquid Neural Networks — Lecture

You reached the end, thanks for reading.
Do these posts resonate with you? Consider supporting the blog by buying me a book.