Eridos builds memory from temporal co‑occurrence — recovering the associations that similarity misses.
Almost every system we use to find and organise information runs on similarity: given a query, return the most alike item. Similarity is genuinely useful but it’s only one way things connect and the field has built almost everything on it alone. Stairs don’t resemble a slip, yet one evokes the other; the smell of sunscreen returns a whole holiday. Those links exist because the things were experienced together, not because they’re alike. Eridos takes that signal — temporal co-occurrence — and builds memory around it: the structure of lived experience reflected in the geometry of an embedding space. Across text, biology and beyond, it recovers associations that similarity misses entirely. The work builds toward systems that understand the reality they occupy.
Eridos runs multiple, related lines of research: associative memory built from experience, computation in optical and crystalline substrates and exploration in multi oscillator reservoir computing. Different layers of the stack but the same instinct: question the assumption everyone else builds on.
The foundation: encoding many weight matrices in one shared holographic volume, and the first hint that adding more sharpens what’s already stored. Simulation.
The mechanism behind it, the geometry that scales it, and a 64-matrix volume read back cleanly. Simulation.
Temporal co-occurrence learning transfers to molecular biology — recovering functional gene relationships (AUC 0.908) where similarity scores near chance.
PAM trained on 9,766 novels discovers what passages do rather than what they’re about — unsupervised hierarchical narrative structure.
The foundational paper. A JEPA-style predictor retrieves true temporal associates 97% of the time, where cosine similarity scores zero.
A 4.2M-parameter reranker lifts HotpotQA Recall@5 by 8.6 points — +28.5 where the dense baseline fails — in under two minutes of training.
A reliability-weighted learning mechanism where plasticity reactivates automatically under distribution shift.
Experimental confirmation of the core mechanism, with boundary conditions identified in tightly-coupled architectures.
A model trained on 10,000 novels discovers what text does rather than what it’s about. Explore the clusters interactively.
In simulation, packing many weight matrices into one shared wave-optical volume — where adding more can sharpen the patterns already stored rather than degrade them.