robotmem introduces a persistent memory system for AI robots that stores and retrieves physical experiences to improve future performance. The system records parameters, trajectories, and outcomes from robot experiments, then uses hybrid search to find the most relevant past experiences for guiding new decisions. A FetchPush experiment demonstrated a 25% success rate improvement, jumping from 42% to 67% using only CPU processing.

Built as an MCP Server with seven APIs, robotmem combines BM25 text search, vector search, and spatial retrieval to understand the structure of robot experiences. The system features context filtering to retrieve only successful experiments and spatial sorting to find nearby scenarios. Each memory includes structured data about parameters, spatial positioning, robot details, and task outcomes. When an episode ends, the system automatically consolidates similar memories and proactively recalls relevant experiences for the next task.

Using SQLite with FTS5 and vec0, robotmem operates entirely on CPU without requiring a GPU. The open-source project includes a web management UI and can be installed via pip. Compared to alternatives like MemoryVLA and Mem0, robotmem uniquely offers structured filtering, spatial retrieval, and physical parameter storage specifically designed for robotic applications. The system addresses a critical challenge in robotics: robots currently repeat experiments from scratch instead of learning from past experiences.