Google drops Always On Memory Agent

This project matters because it argues bigger context windows are not the same thing as usable memory.

• –The README explicitly positions the agent against vector DB + RAG, using an LLM to continuously read, summarize, connect, and rewrite structured memory instead of only retrieving chunks on demand
• –It supports multimodal ingestion from text, images, audio, video, and PDFs, then runs periodic consolidation to produce higher-level insights and links across memories
• –The shipped implementation is tightly coupled to Gemini and Google ADK, so local-model support looks possible in principle but not like a simple config swap
• –Even with 1M-token local models, the architecture still has value because background compression, prioritization, and synthesis solve a different problem than just stuffing more raw context into a prompt