RuleMemory Cloud

▶ Live app: https://rulememory-cloud-516399822756.us-central1.run.app · ▶ Demo video: https://youtu.be/7YjfPy0QL84

A Gemini-powered, MongoDB-backed memory agent that turns contest and launch rules pages into auditable, queryable memory — and runs a multi-step task under user oversight, not just a chat turn. Built for the Google Cloud Rapid Agent Hackathon on the MongoDB partner track, deployed live on Cloud Run.

Try it now

The link above is a live deployment. Click Run agent to watch the full pipeline execute against real Gemini + MongoDB, then ask the memory a question. No setup required.

The multi-step task (under oversight)

ingest a rules page
  → Gemini (Google Gen AI SDK) extracts structured facts: deadlines, rules, eligibility
  → facts written to MongoDB through the official MongoDB MCP server (live tools/call)
  → facts persisted to the durable system of record
  → conflict / supersede: a new fact on the same topic retires the stale one
  → query deadlines expiring within N hours
  → flag assumptions past their stale-after window
  → Gemini grounds a natural-language answer, citing the remembered fact IDs

Every step appends a typed entry to an inspectable transcript, so a human can audit exactly what the agent remembered, changed, and concluded.

Why it is more than a demo

Live partner-MCP integration. Real MongoDB MCP server running as an in-container sidecar with the initialize handshake + session reuse — the agent writes through tools/call, not a mock.
Idempotent + conflict-aware memory. Facts are pinned by entry ID so the MCP write and the durable upsert never collide; contradicting facts are superseded instead of silently duplicated.
A real web console. Status badges, an animated agent plan, grounded answers, a live memory table, and an "ask existing memory" box — the same surface a judge sees in the video.