Comparing Memory Backends (long-context, Mem0, Letta, Graphrag) On Agent Tasks

MemoryArena is a research framework and benchmark for agent memory in interdependent multi-session agentic tasks. It wires pluggable memory backends (long-context, mem0, Letta, Mirix, GraphRAG, MemoRAG, BM25) into task agents and step-based environments for web shopping, travel, search, and formal reasoning.

AI Agents | Python · agent-memory · benchmark