Openwist

Project Cortex is an architecture for machines that have to reason and act in the real world.

Project CortexResearch paper, July 2026
Read PDF

The problem

Robots and vehicles perceive continuously, must remember what they have seen, decide quickly on limited hardware, and keep learning without forgetting. Today those needs are usually met by separate pieces that are hard to inspect and hard to trust.

Our approach

One architecture built from separate parts for each of those jobs, so its behaviour is easier to reason about, test and explain. Cortex combines liquid time-constant networks, state-space layers, Hopfield-style associative memory, and continual learning with elastic weight consolidation and replay.

First results

Our research paper, published in July 2026, reports three of the five experiments we have planned. Each one tests a single component on its own. The paper and its charts use the project's earlier name, NEXUS-1.

Memory: EXP-002

A modern Hopfield memory kept recall at 1.000 up to 800 stored patterns at 128 dimensions with 10 percent noise, while a classical Hopfield memory failed at around 25 patterns. At 64 dimensions, recall at 800 patterns was 0.88. We did not find a capacity limit at 128 dimensions or more. The test used random patterns, so real data may behave differently.

Fast control: EXP-004b

Four small models with matched parameter counts (769 to 827) were compared on two control tasks, using three random seeds. Our fast-reflex model, SpinalLNN, ran in 0.7 ms per step on a CPU.

Pendulum-v1 learning curves for four models
Pendulum-v1. Higher is better. SpinalLNN ends highest, at about -1213 on the rolling 100-episode mean, against -1514 for the GRU and -1677 for the MLP. No model reached the solve threshold.
CartPole-v1 learning curves for four models
CartPole-v1. SpinalLNN finished last of the four, at about 344 against about 499 for the MLP.

The results are mixed. SpinalLNN did best on continuous control and worst on the simpler discrete task.

Continual learning: EXP-005

A network trained on MNIST and then on Fashion-MNIST kept 28.4 percent of its first-task accuracy with no protection, and 98.1 percent with replay and elastic weight consolidation together. Replay alone kept 97.4 percent and elastic weight consolidation alone kept 54.6 percent, so in this test replay did nearly all of the work. Learning the second task was not harmed.

What we do not know yet

Long context on small memory

We also want language models that run in very little memory and can follow very long inputs. A standard transformer keeps a record of every earlier token, so its memory use grows with the length of the text. A state-space layer keeps a fixed-size state instead, so its memory use does not grow with length. That is why Cortex uses state-space layers for long sequences.

1 article

Explore work from other areas