MemRefine: LLM-Guided Compression for Long-Term Agent Memory

In three linesMemRefine compresses LLM agent memory over long-term interactions by using an LLM judge to decide merge, delete, or preserve decisions based on factual content rather than surface similarity. Tested across multiple long-conversation benchmarks, the system meets memory budgets while preserving downstream performance.

Summary generated by Claude — human-verified