LightMem-Eego introduces lightweight streaming multimodal memory for everyday-life assistance

Researchers present LightMem-Ego, a system designed to help personal AI assistants on mobile and wearable devices manage long-term user experiences through visual and audio streams. The architecture continuously captures egocentric data, aligns it on a shared timeline, and organizes it into current, short-term, and long-term memory levels.

• Dynamically routes retrieval to the appropriate memory level based on user queries.
• Generates answers grounded in multimodal evidence from the hierarchical memory structure.
• Supports applications such as object finding, conversation recall, life summarization, routine discovery, and personalized assistance.
• Demonstrated deployment on smartphones and AI glasses.

The system addresses the challenge of answering queries about past experiences by providing a lightweight mechanism for continuously accumulating and retrieving long-term memories.