Google launches Gemini agents with always‑on memory, boosting generative AI capabilities.

Google’s Gemini agents now run an “always‑on” memory layer that lives as a lightweight background process, continuously ingesting, consolidating and querying information without the need for external vector stores or embedding pipelines, the Google Cloud generative‑AI report explains. The architecture assigns each agent a dedicated set‑of‑tools for reading and writing a structured memory store, while an orchestrator routes incoming requests to the appropriate specialist component. By keeping the memory active 24/7, the system mimics the human brain’s sleep‑time replay, linking disparate facts and compressing them into higher‑level insights (Google Cloud Platform, “Always On Memory Agent”). This marks a departure from the “amnesiac” behavior of most LLM‑driven agents, which only process data on demand and discard it afterward.

The ingest stage can handle any of 27 file types—including text, images, audio, video and PDFs—through three entry points: a file‑watcher that picks up drops in an ./inbox folder, a Streamlit dashboard upload button, or a simple HTTP POST to /ingest (Google Cloud Platform). The IngestAgent leverages Gemini 3.1 Flash‑Lite’s multimodal capabilities to extract structured summaries, entity lists, topic tags and an importance score, as illustrated by the example where a news snippet about Anthropic’s Claude usage is turned into a concise, machine‑readable record. This multimodal parsing eliminates the need for separate preprocessing steps that traditional pipelines often require.

Every thirty minutes the ConsolidateAgent awakens to run a “sleep‑like” routine that reviews newly ingested memories, discovers cross‑references, and generates insights that are written back into the memory store. The report shows how the agent links a note about AI‑agent reliability with another about cost‑reduction priorities, surfacing a higher‑level observation that “the bottleneck for next‑gen AI tools is the transition from static RAG to dynamic memory systems.” By continuously stitching together these connections, the system builds a living knowledge graph without the overhead of manually maintained knowledge‑graph databases, a pain point highlighted in the same source.

Querying the memory is handled by a dedicated QueryAgent that reads the consolidated store and returns answers with explicit source citations. In the sample interaction, a user asks “What should I focus on?” and the agent replies with a prioritized list that references specific memory entries (e.g., “Ship the API by March 15 [Memory 2]”). This citation‑rich response format gives developers confidence in the provenance of the information, addressing a common criticism of black‑box LLM outputs. The same mechanism underpins Google’s newly released free Gemini‑powered Data Science Agent on Colab, which uses the always‑on memory to sort and analyze large datasets on‑the‑fly (VentureBeat, “Google launches free Gemini‑powered Data Science Agent”).

Google’s decision to open‑source the Always On Memory Agent further signals a strategic shift away from vector‑database‑centric retrieval. As VentureBeat notes, the move “ditches vector databases for LLM‑driven persistent memory,” positioning Google’s Gemini platform as a more integrated, end‑to‑end solution for enterprise AI workloads. By eliminating the embedding step and keeping memory active, the approach promises lower latency, reduced infrastructure complexity, and a more human‑like ability to synthesize knowledge over time. Analysts will be watching whether this architecture can scale to the petabyte‑level data volumes typical of large enterprises, but the initial rollout demonstrates that Google is betting on dynamic, self‑organizing memory as the next frontier for generative AI agents.