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

2026-02-13 · Source: cs.NE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Predictive Associative Memory (PAM) is a novel neural architecture designed to overcome the limitations of similarity-based memory retrieval by learning associations through temporal co-occurrence. PAM employs an "Inward JEPA" predictor, trained on continuous experience streams, to map a current state to a region in embedding space containing temporally co-occurring past states. This contrasts with traditional "Outward JEPA" predictors that forecast future states. Evaluated on a synthetic benchmark of 20 rooms and 50 objects, PAM achieved an Association Precision@1 of 0.970, meaning its top retrieval was a true temporal associate 97% of the time. It demonstrated Cross-Boundary Recall@20 of 0.421 where cosine similarity scored zero, and an overall discrimination AUC of 0.916 (0.849 for cross-room pairs) compared to cosine's 0.789 (0.503 for cross-room). Ablation studies confirmed the predictor learned genuine temporal structure, not embedding geometry artifacts, and exhibited anchor-specific recall consistent with episodic memory.

Key takeaway

For research scientists developing advanced AI memory systems, this work demonstrates that incorporating temporal co-occurrence as a primary training signal for associative recall is critical. You should consider implementing a Predictive Associative Memory (PAM)-like architecture to enable systems to form and retrieve episodic memories that are not reliant on representational similarity, thereby enhancing capabilities for grounded experience and long-horizon planning in embodied agents.

Key insights

Temporal co-occurrence enables associative memory recall beyond similarity, crucial for embodied AI.

Principles

• Memory recall benefits from repetition.
• Episodic specificity requires both similarity and association.

Method

PAM uses an Inward JEPA predictor, trained with InfoNCE loss on temporal co-occurrence, to map a composite state to a predicted region of associatively reachable past states in an embedding space.

In practice

• Use fixed training pairs for episodic memory consolidation.
• Integrate temporal co-occurrence for cross-modal associations.

Topics

• Predictive Associative Memory
• Temporal Co-occurrence Learning
• Joint-Embedding Predictive Architectures
• Episodic Memory Recall
• Similarity-Based Retrieval

Code references

• EridosAI/PAM-Benchmark

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI — APEX-MEM enhances long-term conversational AI through structured memory, temporal grounding, and intelligent retrieval.
• CogniFold: Always-On Proactive Memory via Cognitive Folding — CogniFold enables proactive AI agents by continuously self-organizing fragmented event streams into evolving cognitive structures and emergent intents.
• The Memory of AI: 01 — Understanding Memory, Tokens and AI Context — Effective AI agents require a four-part memory architecture to move beyond expensive, amnesiac chatbots.
• Associative-State Universal Transformers: Sparse Retrieval Meets Structured Recurrence — Structured recurrent states can be parameter-efficient for language modeling but require explicit sparse retrieval for exact long-range recall.
• A modular architecture for creating multimodal embodied agents with an episodic Knowledge Graph as an explainable and controllable long-term memory — A modular platform uses an event bus and episodic Knowledge Graph for controllable, explainable multimodal embodied agents.
• SubtleMemory: A Benchmark for Fine-Grained Relational Memory Discrimination in Long-Horizon AI Agents — AI agents struggle to discriminate subtle, context-dependent relations (complementary, nuanced, contradictory) among long-term memories.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.NE updates on arXiv.org.