Building grounded memory systems for embodied intelligence

Founded by Jason Dury in Perth, Australia. We're developing Bernard — an embodied AI system designed to be raised through lived experience rather than trained on static datasets. Our research explores developmental cognitive architectures built on Joint Embedding Predictive Architectures (JEPAs), where meaning exists natively in embedding space and intelligence emerges from prediction, association, and grounding in reality. Three published papers span predictive associative memory, association-augmented retrieval, and unsupervised concept discovery — with an interactive demo exploring 10,000 novels analysed by PAM.

PAM Concept Discovery Demo

Explore 10,000 novels analysed by an unsupervised model that learned narrative structure from temporal co-occurrence. Includes AI-powered structural analysis.

Open demo →

Predictive Associative Memory

A JEPA-style predictor trained on temporal co-occurrence retrieves true temporal associations 97% of the time where cosine similarity fails entirely.

Read more →

Association-Augmented Retrieval

A lightweight neural network trained on passage co-occurrence improves multi-hop retrieval by 8.6 points on HotpotQA, with larger gains on the hardest questions.

Read more →

Concept Discovery Through Predictive Associative Memory

PAM trained at scale on 10,000 novels discovers hierarchical narrative concepts without supervision. Forthcoming.

Read more →

Bernard is the long-term research direction our published work is building toward: an embodied cognitive system whose intelligence emerges from lived experience. Two complementary predictors — one facing outward into the world, one facing inward into memory — operate over a shared embedding space to produce the kind of specific, experiential recall that current systems lack.

Read more →
jason@eridos.ai