Reference implementation of "An Algorithm for Routing Vectors in Sequences" (Heinsen, 2022) and "An Algorithm for Routing Capsules in All Domains" (Heinsen, 2019), for composing deep neural networks.

Implementation of deep implicit attention in PyTorch

Physics-inspired transformer modules based on mean-field dynamics of vector-spin models in JAX

This repository contains the official code for Energy Transformer---an efficient Energy-based Transformer variant for graph classification

[ICML 2024] Uniform Memory Retrieval with Larger Capacity for Modern Hopfield Models

Holographic vectors you can compute with. Bind structure, bundle sets, unbind components cross NumPy, PyTorch, and JAX.

Minimum Description Length Hopfield Networks

Metacognitive AI architecture combining Modern Hopfield Networks with entropy-based confusion detection for insight-like problem solving. Implements energy-based search, adaptive beta scaling, and a Cognitive Database integration for analogical reasoning and topology tunneling.

AI agent memory using Modern Hopfield Networks — no LLM calls, no database, one matrix multiply. MCP server for Cursor, Claude Code, and other AI coding agents.

Simulation code for simulations run in my PhD

A non-parametric memory system inspired by hippocampal architecture. Using a three-stage training curriculum, the model learns to store images in a compressed latent space and retrieve them using corrupted or partial "cues".

Entropic associative memory experiments for facial emotion recognition.

Self-organising attractor memory: text encodes to 64×64 CA attractors, recall is reconstructive, topology emerges. Not a vector DB.

心核: Delta Rule 联想记忆 + LoRA，persona 驱动统一分布训练。短期 W 矩阵 + 未来长期 Memory MLP 双层记忆，小 transformer 涌现个性。Not RAG, not longer context.

A brain-inspired cognitive architecture exploring surprise-gated memory, identity protection, and the Titans/MIRAS framework.

hclust_mix is a Python script that allows the identification of attractor states from gene expression matrices using Hopfield neural networks.

Hopfield model for T=0 and finite T.

Persistent associative memory for LLM agents that learns what connects to what through repeated exposure

TensorFlow optimizer with multi-timescale associative memory based on Google's Nested Learning paper

Fault-tolerant associative memory via 3D Neural Cellular Automata. 100% retrieval at N=10,000 patterns, zero catastrophic forgetting, 27K parameters.