Google Cloud Open-Sources an 'Always-On' Memory Agent That Ditches RAG and Embeddings

The reference design runs continuously on Gemini 3.1 Flash-Lite, letting a language model read, reason and write structured memory to a plain SQLite file instead of a vector database.

Jul 18, 2026Source: MarkTechPost

Google Cloud has open-sourced an "Always-On Memory Agent," a reference implementation that proposes a different way to give AI agents memory — one that throws out the vector database and embeddings most of the industry currently relies on. It landed on July 18, built on Google's Agent Development Kit and Gemini 3.1 Flash-Lite.

The idea

Rather than firing one-shot calls, the agent runs 24/7 as a process. A language model continuously reads, thinks and writes structured memory to a plain SQLite file — treating memory as something maintained over time, the way a person consolidates the day's events, instead of retrieved on demand.

No vector database

The pointed omission is the retrieval stack. There is no vector database and no embeddings; the LLM itself decides what is important and records it as structured text. That trades the speed of similarity search for the judgment of a model that has actually read the material — a direct challenge to retrieval-augmented generation orthodoxy.

How the agents divide the work

An orchestrator routes work to three sub-agents: IngestAgent extracts summaries, entities, topics and importance; ConsolidateAgent synthesizes links across memories every 30 minutes; and QueryAgent answers questions with citations back to source. The system ingests 27 file types across five categories — text, images, audio, video and documents.

Why Flash-Lite

The choice of Gemini 3.1 Flash-Lite is the enabling bet. Continuous background cognition only pencils out if the model is cheap and fast enough to run all the time. As lightweight models improve, always-on consolidation becomes affordable — and Google, conveniently, sells exactly that class of model. It is worth stressing that this is a reference design, not a product; Google is publishing a pattern others can copy, which is precisely how architectural defaults shift. If the approach catches on, a swath of vector-database and embeddings tooling built for one-shot retrieval starts to look optional.

Why It Matters

Vector databases and RAG have been the default memory architecture for AI agents and a whole layer of startups; a credible, Google-blessed pattern that replaces them with continuous LLM consolidation challenges that consensus and points toward cheap models doing always-on background cognition rather than expensive one-shot retrieval. Even as a reference design, it nudges the agent-infrastructure stack — and validates a market for the low-cost models Google itself ships.

Filed underGoogle AI AI Agents AI Tools #Gemini #Google #RAG #memory