Hybrid Associative Memories

Summary

Hybrid Associative Memories are the selective-memory layers introduced by Hybrid Associative Memories. A HAM layer keeps a recurrent state for predictable context and writes only hard-to-predict tokens into a KV-cache scratchpad.

Role In The Wiki

Use this page as the object card for HAM. The source page carries the evidence, limitations, and agenda mapping.

For the foundation time-series agenda, HAM is upstream architecture evidence for data-dependent retained-history budgets. It suggests a possible split between compressed recurrent state for ordinary spans and explicit memory for rare or high-value spans, but it does not directly evaluate numeric time series, event streams, or action-conditioned trajectories.

Evidence

• Hybrid Associative Memories

Official Artifacts

• Preprint: arXiv 2603.22325
• Official Zyphra page: Hybrid Associative Memories
• Related Zyphra cookbook: Zyphra/zcookbook
• Official dedicated code/model release: none found at ingest time.
• Local X provenance snapshot: papers/hybrid-associative-memories-2026/x_discussion_ham_in_oryx_thread_2067675034878193787.md

Related Pages

• Extra-Long Context For Time Series
• Efficient Recurrent Sequence Models
• Streaming Latent-State Updates
• Looped Transformers And Test-Time Memory
• Time-Series Scaling And Efficiency
• Hierarchical Modeling with a Fixed FLOPs Budget
• Gated DeltaNet
• Oryx