MemRefine Uses LLM Judge to Compress Agent Memory Within Fixed Budgets

MemRefine is a framework for storage-budgeted long-term agent memory that uses surface similarity only to propose candidate pairs, then delegates delete, merge, or preserve decisions to an LLM judge based on factual content. It targets resource-constrained deployments where unbounded memory growth degrades retrieval quality.