Google Cloud Memory Always On Agent Replaces RAG and LLM Continuous Integration Embedding in Gemini 3.1 Flash-Lite
Most AI agents are forgettable. They process the request, respond to it, and send the context. Google Cloud's generative-ai The repository now ships with a sample that addresses this directly. This is it Always On Memory Agentreference usage that manages memory as an active process.
Always On Memory Agent
Basically, the project is a lightweight non-stop backend agent. It works 24/7 as an ongoing process, not a one-shot call. Built with Google ADK (Agent Development Kit) and Gemini 3.1 Flash-Lite. Notably, it does not use a vector database and there is no embedding. Instead, the LLM reads, thinks, and writes structured memory to it SQLite. The choice of model targets low latency and low cost for ongoing background work.
How It Works: Import, Merge, Query
Architecturally, the orchestrator routes all requests to one of three sub-agents. Each subagent has its own tools for reading or writing the memory store.
First, i IngestAgent handles incoming content. It uses Gemini's multitasking capabilities to generate a summary, businesses, topics, and a critical score. That structured record then goes into memories the table.
Next, i ConsolidateAgent it works on a timer, every 30 minutes automatically. Like sleep cycles, it updates disorganized memories and finds connections between them. It then writes a combined summary, one key insight, and that connection to the database. As a result, the agent constructs a new understanding while passively, without being informed.
Finally, i QueryAgent he answers the questions. It reads all the memory and configuration information, and compiles the response. Importantly, it cites the memory IDs it used as sources.
