MemoryOS

Why we included this project

Agents that keep forgetting what happened two turns ago are a familiar pain, and MemoryOS treats that problem the way an OS treats RAM and disk: memory moves through storage, updating, retrieval, and generation stages so the agent recalls the right past detail when it matters. It reports strong results on the LoCoMo long-term memory benchmark, roughly 49% higher F1 and 46% higher BLEU-1 over baselines, and the paper was accepted as an EMNLP 2025 oral. Integration is practical: it ships as Python packages plus an MCP server, so you can give existing clients like Claude Desktop, Cline, or Cursor durable memory without building a custom stack. If you're comparing ways to add long-term memory to a personal assistant rather than stitching together a vector store, this is a useful reference to work from.