Arxis Package

pama-cache

Learned KV-cache allocation for long-context LLMs.

pama-cache replaces static eviction methods with a learned policy that allocates information across working, compressed, and associative memory tiers.

  • Task-adaptive retention and eviction behavior
  • Rate-distortion optimization under bit budgets
  • Dynamic tier budgeting as context grows
  • PyPI-ready package for drop-in integration

pip install pama-cache