Predictive Associative Memory: Retrieval Beyond Similarity Through Temporal Co-occurrence

We propose Predictive Associative Memory (PAM), an architecture in which a JEPA-style predictor, trained on temporal co-occurrence within a continuous experience stream, learns to navigate the associative structure of an embedding space. Where similarity-based retrieval (RAG, Hopfield networks) assumes useful memories are representationally close to the query, PAM retrieves associations that cross representational boundaries — states linked by having been experienced together, not by what they look like. We introduce an Inward JEPA that operates over stored experience (predicting associatively reachable past states) as the complement to the standard Outward JEPA that operates over incoming sensory data. On a synthetic benchmark, the predictor achieves Association Precision@1 = 0.970, cross-boundary Recall@20 = 0.421 (cosine similarity: 0.000), and discrimination AUC = 0.916 (0.849 on cross-room pairs where cosine is at chance). A temporal shuffle ablation collapses cross-boundary recall by 90%, confirming the signal is genuine temporal structure. Code: github.com/EridosAI/PAM-Benchmark