WarpState: 2B-param architecture swaps growing KV cache for fixed-size dual associative memory

zemondza · reddit · 2026-09-21

An experimental 2B-parameter architecture, WarpState, replaces ever-growing token-level KV states with a compact recurrent memory: two associative banks per head (fast/slow decays, β≈0.90/0.99) updated by bounded normalized EMA writes, plus exact causal attention within 128-token chunks. Streaming state stays 49 MiB regardless of stream length — though the author stresses stream length ≠ effective memory horizon.

Related event: WarpState: Two-Timescale Memory to Replace KV Cache(2 posts)→

Original post →

More from Research

Research channel →