Remembering Distinct Items, Not Tokens: A Learnable Dirichlet-Process Cache Between State-Space Models and Attention

What the math gives to ML

The paper provides a concrete online clustering operator that turns a recurrent model's unbounded token history into a cache whose capacity scales with distinct key patterns rather than sequence length. Its transferable asset is the hard novelty decision: cosine distance to the nearest occupied slot determines whether to allocate memory or merge with an existing prototype, giving an explicit memory-growth control knob instead of heuristic eviction. The most promising integration is a write-side cache after an SSM or recurrent backbone, retaining representative key-value pairs and using the sparse table for associative recall. This should first be tested against full attention and fixed-budget eviction on redundant streams, measuring recall, cache size, and inference cost at equal sequence length.