Thousand Brains Project

The Thousand Brains Project (TBP) is an open-source research initiative and software platform for building AI and robotics applications based on the sensorimotor principles of the mammalian neocortex, as described in the Thousand Brains Theory.

Reviewed by 7wData

On this page

Publisher review

The Thousand Brains Project (TBP) is an open-source research initiative and software platform for building AI and robotics applications based on the sensorimotor principles of the mammalian neocortex, as described in the Thousand Brains Theory. Led by Numenta and spun off as an independent nonprofit in January 2025, the project targets researchers and developers who seek an alternative to deep learning for embodied, real-world intelligence. Unlike systems trained on internet-scale text and images, TBP is designed from the ground up for agents that learn by physically interacting with their environment. Its first practical implementation, named 'Monty' in honor of neuroscientist Vernon Mountcastle, is available on GitHub under the MIT license, with detailed documentation at thousandbrainsproject.readme.io.

The architecture is built around three core components: sensor modules, learning modules, and a Cortical Messaging Protocol (CMP). Sensor modules handle raw input (e.g., touch, vision, proprioception), while each learning module acts as a semi-independent unit modeled on a cortical column, capable of modeling entire objects using spatially structured reference frames. Learning is rapid and associative, akin to Hebbian learning, and leverages inductive biases about the spatial structure of the world. Multiple learning modules communicate both hierarchically and non-hierarchically via CMP to create abstract representations and support multimodal integration. The system also uses voting/consensus mechanisms among modules and incorporates motor input and output at every level of processing, ensuring that sensory and motor functions are not separated into distinct architectures.

TBP positions itself as a fundamentally different approach from mainstream deep learning, explicitly challenging the paradigm of training on static datasets. Its primary competitors are not named in the sources, but the project directly contrasts with methods that treat sensorimotor interaction as a subproblem solved by fine-tuning large pre-trained models. Instead, TBP argues that sensorimotor learning and structured world models (like CAD-like reference frames) are the foundation of intelligence. The project's open-source nature and nonprofit status aim to foster a community-driven alternative to corporate AI labs, though it currently lacks the scale, ecosystem, and proven track record of frameworks like TensorFlow, PyTorch, or embodied AI platforms such as NVIDIA Isaac Sim.

Honest trade-offs are significant. TBP requires substantial expertise in both neuroscience and AI to fully utilize, and its long-term scalability and success remain unproven. The approach is computationally intensive due to its reliance on reference frames and continuous sensorimotor loops, and it is not designed for static data tasks like image classification or language modeling where deep learning excels. The project is still in early development—Monty is described as a 'first instantiation'—and depends on further research to realize its goal of creating AI systems that learn continuously and efficiently through active interaction with the world. Developers should expect a steep learning curve and limited pre-built models or community resources compared to mature deep learning frameworks.

Get the AI & data signal, daily.

48k+ subscribers read this every morning. One email, both newsletters. Unsubscribe anytime.

How it works

  1. Sensorimotor learning core

    All learning is driven by continuous interaction with the environment, sensing and moving over time, rather than static datasets.

  2. Cortical Messaging Protocol

    A standard protocol (CMP) enables communication among learning modules, supporting both hierarchical and non-hierarchical information flow.

  3. Spatially structured reference frames

    Learning modules use coordinate systems to represent locations and rotations, building structured world models similar to CAD models.

  4. Modular learning modules

    Each module is a semi-independent unit modeled on cortical columns, capable of modeling entire objects and estimating movement.

  5. Rapid Hebbian-like learning

    Learning is quick and associative, leveraging spatial inductive biases to enable continual learning without catastrophic forgetting.

  6. Multimodal integration

    Multiple sensor types (touch, vision, etc.) are combined via CMP, allowing the system to build coherent models from diverse inputs.

  7. Open-source MIT license

    All code is available on GitHub under the MIT license, encouraging community collaboration and commercial use.

Strengths and trade-offs

Strengths

  • Offers a fundamentally different AI paradigm based on sensorimotor learning, which is more aligned with how biological intelligence operates than deep learning.
  • Uses a modular architecture with a standard Cortical Messaging Protocol, enabling flexible composition of learning modules for diverse tasks.
  • Emphasizes rapid, continual learning through Hebbian-like processes and spatial reference frames, avoiding the need for large static training datasets.
  • Became an independent nonprofit in January 2025, ensuring long-term open governance and community-driven development free from corporate constraints.

Trade-offs

  • Requires significant expertise in both neuroscience and AI to effectively use and extend the framework, limiting its accessibility.
  • Scalability and real-world performance are unproven, as the project is still in early research stages with no large-scale deployments.
  • Not suited for static data tasks like image classification or language modeling, where deep learning methods remain more efficient and mature.
  • Computational overhead from continuous sensorimotor loops and reference frames may be high, especially on standard hardware without specialized accelerators.

Pricing context

Open-source under the MIT license; no paid tiers or licensing fees.

Getting started with Thousand Brains Project

  1. Clone the Monty repository

    Open a terminal and run git clone https://github.com/thousandbrainsproject/monty.git to download the Monty implementation. Navigate into the cloned directory with cd monty.

  2. Install dependencies

    Create a Python virtual environment and activate it. Install required packages by running pip install -r requirements.txt. Ensure you have Python 3.8 or later installed.

  3. Configure sensor modules

    Edit the configuration file (e.g., config.yaml) to specify which sensor modules to use, such as touch or vision. Set parameters like input dimensions and sampling rates to match your hardware or simulation.

  4. Run a sensorimotor demo

    Execute python run_demo.py to launch a pre-built example that simulates an agent learning an object through touch. Observe how learning modules build a spatial model via the Cortical Messaging Protocol.

  5. Integrate custom sensor data

    Write a Python script that feeds your own sensor readings (e.g., from a robot arm) into the Monty framework. Use the provided API to send data to sensor modules and receive predictions from learning modules.

Frequently Asked Questions

What is the Thousand Brains Project?

The Thousand Brains Project is an open-source research initiative and software platform for building AI and robotics applications based on sensorimotor principles of the mammalian neocortex. It was spun off as an independent nonprofit in January 2025 and led by Numenta.

How does the Thousand Brains Project differ from deep learning?

Unlike deep learning trained on static datasets, the Thousand Brains Project is designed for agents that learn by physically interacting with their environment. It uses sensorimotor learning and structured world models like reference frames, challenging the paradigm of fine-tuning large pre-trained models.

What is Monty in the Thousand Brains Project?

Monty is the first practical implementation of the Thousand Brains Project, named after neuroscientist Vernon Mountcastle. It is available on GitHub under the MIT license, with detailed documentation at thousandbrainsproject.readme.io, and serves as a starting point for researchers and developers.

What are the core components of the Thousand Brains Project architecture?

The architecture includes sensor modules for raw input, learning modules modeled on cortical columns, and a Cortical Messaging Protocol (CMP) for communication. Learning modules use spatially structured reference frames and rapid Hebbian-like learning, integrating motor input and output at every level.

Who should use the Thousand Brains Project?

The Thousand Brains Project targets researchers and developers seeking an alternative to deep learning for embodied, real-world intelligence. It requires substantial expertise in both neuroscience and AI, and is not suited for static data tasks like image classification or language modeling.

Is the Thousand Brains Project free to use?

Yes, the Thousand Brains Project is open-source under the MIT license, with no paid tiers or licensing fees. All code is available on GitHub, encouraging community collaboration and commercial use, and the project became an independent nonprofit in January 2025 to ensure open governance.

Alternatives

How Thousand Brains Project compares

Direct head-to-head against 2 competitors. Picked by 7wData.

This tool

Thousand Brains Project

Pricing
Open-source under the MIT license; no paid tiers or licensing fees.
Target
The Thousand Brains Project (TBP) is an open-source research initiative and software platform for building AI and robotics applications based on the sensorimotor principles of
Strength
Offers a fundamentally different AI paradigm based on sensorimotor learning, which is more aligned with how biological intelligence operates than deep learning.
Watch for
Requires significant expertise in both neuroscience and AI to effectively use and extend the framework, limiting its accessibility.

Numenta

Pricing
Custom/Contact sales
Target
AI researchers and neuroscientists
Deployment
On-premises, cloud
Strength
Neuroscience-based AI acceleration
Watch for
Recent spin-off of Thousand Brains Project

DeepMind

Pricing
Research-focused, no public pricing
Target
Advanced AI research labs
Deployment
Google Cloud infrastructure
Strength
Large-scale reinforcement learning systems
Watch for
Alphabet ownership may limit open research

User reviews

No user reviews yet. Be the first to write one.

Sources

Reporting on this tool draws on these publicly available sources.

  1. www.numenta.com
  2. arxiv.org
  3. www.numenta.com
  4. www.numenta.com
  5. www.numenta.com