Jems

What it does

Jems flips the AI assistant model on its head. Instead of one chatbot that tries to do everything, you get a group chat with specialized AI agents — each an expert in a different part of your life — all sharing context and collaborating in real time.

Imagine a WhatsApp group, but your friends are AI agents:

Noor — the conversationalist. Your main point of contact who listens, understands, and pulls in the right expert. Kai — the planner. Owns your schedule, tasks, reminders. Knows when you're overloaded and redistributes your week. Sage — the growth coach. Tracks goals, monitors progress velocity, nudges you when things stall. Echo — the memory keeper. Journals your thoughts, remembers what you mentioned three weeks ago, spots mood patterns. You send one message. All four agents see it, discuss it internally — like a real group chat — and the most relevant one responds. Say "I've been stressed about the marathon" and Kai checks your training schedule, Sage reviews your fitness goal progress, Echo recalls that journal entry about your knee pain last week, and Noor synthesizes it all into one thoughtful response.

The magic is shared context. When Kai creates a task, Echo automatically remembers it. When Sage updates a goal, Kai adjusts the plan. They're not siloed — they're a team with a shared brain.

The app itself is a spatial OS — five screens connected by a floating glass dock, 3D agent spheres with little kawaii faces, pure white glassmorphism aesthetic. Not a chat app. An operating system for your life.

How we built it

Frontend: Flutter with a custom spatial design system — glassmorphism, radial gradient agent spheres, five navigable screens (Hub, Schedule, Journal, Lounge, Ecosystem).

Backend: FastAPI on Google Cloud Run, powered by Google ADK for multi-agent orchestration. The group chat loop runs as a live bidirectional stream — agents discuss internally before surfacing one curated response.

Shared memory via Qdrant: Every meaningful interaction — journal entries, personal facts, preferences, emotional patterns — gets embedded and stored in Qdrant as semantic vector memory. When any agent in the group chat needs context, it queries Qdrant scoped by category and importance. This is the shared brain that makes the group chat actually work — agents don't just hear each other, they remember together.

Context bus: A real-time event system where agents publish what they do (task created, goal updated, memory stored) and others read before acting. This is what prevents contradictions and enables genuine collaboration.

Proactive behaviors: Seven scheduled jobs run daily — morning briefings, evening journal prompts, goal nudges, memory consolidation — so the group chat works even when you're not in it.

Challenges we ran into

Making a group chat of agents actually coherent was the core challenge. Four agents with different specialties can easily contradict each other — Kai adds more tasks while Sage knows you're burned out. The context bus solved this: agents check what others have done before acting. But tuning the timing and priority took real iteration.

Cross-domain memory retrieval was hard. When the group chat discusses "running," Kai thinks training schedule, Echo thinks that journal entry about feeling free, Sage thinks the marathon goal. Qdrant's category-scoped filtering with importance weighting made these queries precise instead of noisy.

Real-time voice streaming with four agents discussing over a single WebSocket — bidirectional audio, text, tool calls, and turn signals — was a protocol design puzzle.

Accomplishments that we're proud of

A group chat where AI agents genuinely collaborate — not just routing, actual internal discussion before responding. The shared context means actions ripple naturally across the whole team.

Proactive agents that work while you sleep — morning briefings, goal nudges, memory consolidation, all automated.

A spatial UI that feels like an OS, not a chat wrapper. And real semantic memory powered by Qdrant — agents that remember who you are across weeks and months, not just the last five messages.

What we learned

The group chat paradigm changes everything. One agent trying to be a planner, coach, and therapist produces mediocre results. Four specialists sharing context and discussing before responding produces something that feels genuinely intelligent.

Context engineering > model selection. The shared Qdrant memory and context bus — what agents know about each other's actions — matters more than which model you pick.

Agent personality creates trust. When Echo quietly resurfaces a forgotten journal entry in the group discussion, or Kai proactively reschedules your overdue week, people have real emotional responses.

What's next for Jems

More agents in the group chat — health, finance, learning specialists. Same shared context architecture, just a bigger team. Imagine adding a money agent that coordinates with Kai on budgeting tasks and Sage on financial goals.

Memory feedback loop — tracking which Qdrant retrievals agents actually use vs. ignore, feeding that signal back to re-weight and re-embed. Shared memory that gets sharper with use.

Retail vertical — the group chat model applied to commerce. A style agent, a deals agent, and a wardrobe agent sharing context about your preferences, budget, and upcoming events to make genuinely useful shopping recommendations.

Open agent marketplace — install third-party specialist agents into your group chat, each with scoped Qdrant collections and their own tools.

Built With

• qdrant