Google Cloud Ships Always-On Memory Agent for LLMs

This agent treats memory as a continuous background process that runs 24/7, rather than a single-use call. It is built using Google's Agent Development Kit (ADK) and leverages the Gemini 3.1 Flash-Lite model, notably omitting the need for vector databases and embeddings.

Instead, it utilizes an LLM to read, process, and write structured memory into an SQLite database.

The architecture of the Always-On Memory Agent involves an orchestrator that directs incoming requests to one of three specialized sub-agents. The IngestAgent is responsible for processing new information, employing Gemini's multimodal capabilities to extract summaries, entities, topics, and an importance score, which are then stored in a memories table.

The ConsolidateAgent operates on a default 30-minute timer, reviewing unconsolidated memories to identify connections and synthesize new understanding. This process involves writing a synthesized summary, a key insight, and identified connections back into the database, allowing the agent to build knowledge even when idle and without direct prompting.

Finally, the QueryAgent handles user questions by accessing all stored memories and consolidation insights. It synthesizes a response and crucially provides citations for the memory IDs used in its answer. The choice of Gemini 3.1 Flash-Lite is intended to support low latency and cost-effectiveness for continuous background operations. This approach represents a shift from traditional RAG (Retrieval-Augmented Generation) methods, aiming for more persistent and integrated memory management within AI agents.