Google Cloud Releases Always-On Memory Agent Powered by Gemini 3.1 Flash-Lite
Replace traditional Retrieval-Augmented Generation (RAG) databases with a background agent that consolidates memory into SQLite. Running on Gemini 3.1 Flash-Lite and Google Agent Development Kit (ADK), it actively links and synthesizes details.
Impact: High
Why it matters
You can build a low-cost, zero-embedding memory loop that processes documents, audio, and images continuously without vector databases.
TL;DR
- 01Replaces vector embeddings and traditional RAG with continuous LLM consolidation on SQLite.
- 02Ingests 27 distinct file formats including markdown, CSV, logs, images, and audio tracks.
- 03Runs on Google Agent Development Kit and highly optimized Gemini 3.1 Flash-Lite to keep computing costs negligible.
Key facts
- Supported file formats
- 27 categories
- Consolidation interval
- 30 minutes (default)
- Query context depth limit
- 50 recent memories
Architecture: The Three Sub-Agents
Google Cloud's approach structures memory into three modular roles: 1. IngestAgent: Automatically parses 27 supported file types (covering raw logs, markdown, and audio) dropped in the ./inbox folder. It evaluates file contents and logs metadata directly to SQLite tables. 2. ConsolidateAgent: Acts as a background sleep-cycle processor. Every 30 minutes, it analyzes non-linked database rows, finds connections, and updates the knowledge store. 3. QueryAgent: Evaluates active user questions by reading up to 50 recent memories and synthesizes structured citations.
Running the Local API
You can launch the background agent with direct command-line arguments to customize port configurations and directory paths:
python agent.py --watch ./docs --port 9000 --consolidate-every 15Once operational, send raw text prompts via standard HTTP requests:
curl -X POST http://localhost:8888/ingest \
-d '{"text": "AI agents are the future", "source": "article"}'Try it in 2 minutes
python agent.py --watch ./docs --port 9000 --consolidate-every 15bash
✓ When to use
- Perfect for building personal knowledge bases, local log collectors, and autonomous researchers.
- Ideal when deploying zero-cost, serverless AI tasks where vector databases are too heavy or expensive.
✕ When NOT to use
- Do not use if your workloads require high-velocity, sub-millisecond retrieval of unstructured millions-node vector indexes.
- Avoid if you require strict real-time transactional synchronization of user context.
What to do today
- Set your GOOGLE_API_KEY environment variable locally.
- Clone the repository and run the agent to auto-parse file updates inside your workspace folders.
- Configure curl webhooks to stream live logs or notes directly to your local database backend.
Sources