LLM Efficiency and Memory Architectures
Key Questions
What speedup does RTPurbo achieve for long-context LLMs?
RTPurbo converts full attention to sparse attention in 100 steps, delivering a 9.4x prefill speedup at 1M context length with near-lossless performance.
How does δ-mem improve memory for long-term agents?
δ-mem introduces a compact 8x8 associative memory state with delta-rule updates. It outperforms baselines without context rot and aligns with sparsity trends like MLA.
What overall trend do these efficiency methods support?
Both RTPurbo and δ-mem advance sparse and compact memory architectures. They target practical deployment of efficient LLMs and agents.
RTPurbo converts full attention to sparse in 100 steps with 9.4x prefill speedup at 1M context, near-lossless. δ-mem introduces compact 8x8 associative memory state with delta-rule updates for long-term agents, outperforming baselines without context rot. Aligns with sparsity trends like MLA.
Sources (2)
Updated May 23, 2026