agent-memory-benchmarks

An open benchmark and reference for evaluating AI agent memory systems.

Every system is measured the way an agent uses memory: it ingests each conversation through its own extractor, an agent searches it through the system's own tools and answers, and an LLM judge marks every answer correct or not. Judge accuracy is the score the summary ranks by, with memory lift beside it — the accuracy the memory added over what the agent answers without the evidence. The attribution table then says who lost each question: the memory, by missing an evidence session, or the agent, by fumbling evidence it was handed. Retrieval precision and recall are not scored.