GroupGPT: A Production-Grade, End-to-End Encrypted AI Group Chat Platform with Structured Knowledge Memory and Hybrid Edge-Cloud Inference

1 · Introduction

The landscape of LLM-based assistants has been dominated by single-user dialogue systems. Extending these to group chat settings—where multiple participants interact simultaneously, conversations shift topic rapidly, and sensitive personal information is regularly shared—introduces a fundamentally different set of requirements.

Related work has addressed group chat AI through frameworks that focus primarily on when the AI should intervene, but leave critical production concerns unaddressed: How does the AI accumulate room-specific knowledge across sessions? How do users protect sensitive information from the platform operator itself? How does the system remain affordable as group size and message volume grow?

GroupGPT addresses all of these in a single, coherent, production-deployed system. Our key contributions are:

• End-to-end encryption with cryptographic AI participation. Private rooms use ECDH P-256 + AES-256-GCM. The AI holds its own ECDH keypair and participates as a cryptographic peer — enabling encrypted AI responses without exposing plaintext to the database.
• Neuron/Brain system: typed, room-scoped knowledge memory. Each room maintains a graph of typed knowledge nodes (skill, knowledge, process, memory) injected into every AI request — persistent context without growing token cost.
• Hybrid edge-cloud inference. On-device WebLLM handles offline and privacy-sensitive queries; cloud inference (Gemini 2.5 Flash via OpenRouter) handles streaming complex responses. The switch is transparent to users.
• Rich multimodal command ecosystem. Seven ! commands — search!, browse!, learn!, code!, image generation, auto URL preview, and report! — plus WebRTC P2P voice chat with Whisper transcription.
• Dual-context architecture with database-level isolation. World chat (public) and private rooms are backed by completely separate SQLite databases.
• Multilingual fan-out translation. Gemini 2.0 Flash fans translated messages only to users with translator mode on — zero overhead for monolingual groups.

2 · Related Work

2.1 Multi-User Group Chat AI Frameworks

MUCA [4] formalized the "3W" dimensions for group chat agents and introduced fixed-interval LLM evaluation. HUMA [5] extended this with human-like behavioral timing. MAP [6] explored multi-agent personalization. Of these, only MUCA (January 2024) predates GroupGPT's public launch on April 4, 2025; HUMA and MAP appeared as concurrent work later in 2025. These frameworks share a common limitation: they are research prototypes that do not address E2EE, persistent structured memory, or production-scale token economics.

A concurrent paper (Shen et al., 2026) [7] proposes an edge-cloud architecture decoupling intervention timing from response generation, introducing the MUIR benchmark. GroupGPT is complementary: where Shen et al. focus on when to intervene and privacy sanitization, GroupGPT addresses what the AI knows (Neurons), cryptographic privacy (full E2EE), and full production deployment.

2.2 Memory-Augmented LLMs

MemGPT [8] introduced hierarchical memory management for LLMs. RAG systems [9] retrieve relevant documents at inference time. GroupGPT's Neuron system differs from both: rather than automatic extraction or document retrieval, Neurons are explicitly curated by room members, giving users full authorship over what the AI knows and ensuring no unintended information leaks between rooms.

2.3 End-to-End Encryption in AI Systems

Signal [10] demonstrated E2EE in messaging. Extending E2EE to AI-assisted conversations — where the AI must read encrypted messages to respond — is an open problem that, to our knowledge, no prior group chat AI system addresses. GroupGPT resolves this by issuing the AI assistant its own ECDH keypair so it participates as a cryptographic peer, identical to human users.

2.4 Hybrid Edge-Cloud Inference

WebLLM [11] enabled in-browser LLM inference via WebGPU. Prior group chat systems do not integrate on-device inference. GroupGPT uses WebLLM as a genuine fallback for offline use and as a privacy-preserving alternative for queries users prefer not to send to cloud APIs.

3 · System Architecture

3.1 High-Level Overview

GroupGPT consists of two independent services: a Frontend (React 18 + Vite + TypeScript, Chakra UI) and a Backend (Node.js + Express + Socket.io, dual Prisma-managed SQLite databases).

3.2 Message Routing Pipeline

Every incoming message passes through a multi-stage decision tree before a response is emitted: security checks → rate limiting → optional E2EE decryption → token budget check → command detection → Neuron-augmented AI streaming → optional re-encryption and translation fan-out.

3.3 World Chat vs. Private Rooms: Dual-Database Isolation

GroupGPT enforces strict isolation at every layer. World chat (world.db → WorldMessage, no roomId) and private rooms (dev.db → RoomMessage, indexed by (roomId, timestamp)) are backed by completely separate SQLite files. The critical invariant is that session.roomId is mutated in-place on join_room; creating a new session object was the historical root cause of world-chat history bleeding into private rooms.

4 · Neuron / Brain System

4.1 Motivation and Design

Prior group chat AI systems use flat chat history as their only source of context. This is expensive (token cost grows linearly with history length), fragile (key facts scroll out of context), and nonspecific (the AI must rediscover room norms from history on every session). GroupGPT's Neuron system solves all three by treating room context the way Neuron Surgery [13] treats model cognition: distinguishing raw knowledge from filtered experience from applied wisdom. Flat chat history is raw knowledge — a growing ledger the AI must re-parse on every request. Neurons are curated wisdom: room members act as human mentors, explicitly shaping what the AI should know, how it should behave, and what constraints apply, rather than leaving it to infer context from noise.

Room members curate a graph of typed knowledge nodes, stored in dev.db and injected into the AI's system prompt alongside a sliding window of recent messages:

4.2 Token Efficiency

Figure 6 compares token usage across three context strategies for an active room with 90 days of history.

4.3 Neuron Portability

Neurons can be exported as .neurons files and imported into other rooms, enabling teams to share AI configuration across projects. The frontend provides a canvas-based graph visualization (NeuronPanel.tsx) for visual editing.

5 · End-to-End Encryption with AI Participation

5.1 Cryptographic Protocol

GroupGPT implements E2EE where the AI assistant participates as a full cryptographic peer. The scheme uses ECDH P-256 key agreement to derive per-room shared secrets, then wraps an AES-256-GCM room key for every participant — including Cortez (the AI), whose ECDH keypair is loaded at server start from the CORTEZ_PRIVATE_KEY_JWK environment variable.

The key exchange sequence:

1. Each user generates a permanent ECDH keypair stored in localStorage and uploads the public key to /api/e2ee/public-key.
2. The room creator generates an AES-256-GCM room key and ECDH-wraps it for every member and Cortez, storing all bundles at /api/e2ee/room-key.
3. On join, each user unwraps their bundle to recover the room key.
4. The server unwraps Cortez's bundle at ChatController initialization and caches it in cortezRoomKeys: Map<roomId, CryptoKey>.
5. All subsequent messages are AES-GCM encrypted; the server decrypts transiently in memory for Cortez, never writing plaintext to the database.

5.2 Security Properties

6 · Multimodal Command Ecosystem

GroupGPT provides a natural-language ! command interface that extends AI capabilities with live data and execution tools. Command detection in commands/detect.ts runs before the LLM streaming path, bypassing the AI entirely for commands that don't require it.

7 · Voice Chat and Speech-to-Text Pipeline

7.1 WebRTC Peer-to-Peer Architecture

GroupGPT supports real-time P2P audio calls within any room. The server acts as a pure signaling relay — no audio traverses the server during live calls. ICE negotiation uses Google STUN and openrelay.metered.ca TURN servers. Clients exchange voice_offer, voice_answer, and voice_ice_candidate Socket.io events, then establish a direct WebRTC audio track.

7.2 Voice Message → AI Pipeline

Beyond live calls, users can send voice messages that route through the full AI pipeline:

1. Browser records audio and emits a voice_message Socket.io event with a raw buffer + MIME type.
2. Server converts the buffer to WAV via ffmpeg-static.
3. WAV is sent to OpenAI Whisper API → returns transcript text.
4. Transcript is routed through the standard message.handler.ts pipeline — including command detection and AI streaming.

A voice message saying "search! latest Node.js release" will trigger the Tavily search command and return results to the entire group, just as a typed message would.

8 · Hybrid Edge-Cloud Inference

8.1 Inference Routing

GroupGPT supports two inference backends, selected transparently based on availability and user preference:

8.2 Multilingual Translation Fan-Out

A background TranslationService (Gemini 2.0 Flash) fans translated message copies only to users who have translator mode enabled and whose configured language differs from the message language. A 10-person English-only room incurs zero translation calls per message. A mixed 10-person room with 3 users needing translation incurs 3 calls — never 10.

9 · Security Subsystem

GroupGPT employs a five-layer concentric security model from network edge to AI response.

10 · Authentication and Subscription

10.1 Authentication Paths

Three authentication paths converge onto JWT issuance by AuthManager: email/password (bcrypt), Google OAuth via Passport.js (/auth/callback), and Firebase (frontend Google sign-in fallback). All paths result in a JWT that is validated on Socket.io handshake.

10.2 Token Budget Model

Before each AI call: check availableTokens >= 500. After: UserManager.addTokenUsage(tokensUsed). Stripe manages billing; webhooks update UserManager state.

11 · Evaluation

11.1 Response Quality (N = 300)

AI responses were evaluated on four dimensions using LLM-as-judge methodology [12] on a stratified sample of 300 responses across 50 private room sessions spanning engineering, creative writing, and support use cases.

11.2 Command Latency Profile

12 · Discussion

12.1 Comparison with Shen et al. (2026)

12.2 Neuron System vs. Automatic Memory Extraction

Automatic memory extraction (MemGPT [8], RAG [9]) requires no user effort but introduces risk: automatically extracted "memories" can silently include sensitive or incorrect information, and the AI's knowledge base grows opaque over time. GroupGPT's Neurons require explicit curation — a higher bar — but give room members full authorship over what the AI knows. Neuron Surgery [13] frames this as the difference between knowledge and wisdom: an AI with automatically harvested facts may be well-informed but still miss the mark, while one whose context has been deliberately shaped by the people it serves is far more likely to produce outputs that feel right to them. The visual Neuron graph editor in NeuronPanel.tsx lowers the curation effort substantially, and the .neurons export format enables reuse across rooms.

12.3 Limitations and Future Work

13 · Conclusion

GroupGPT is a production-grade AI group chat platform that advances the state of the art across five dimensions simultaneously: structured persistent knowledge memory (Neurons), cryptographic end-to-end encryption with AI participation, hybrid edge-cloud inference for privacy and offline support, a rich multimodal ! command ecosystem spanning voice, code, research, and image generation, and a dual-database architecture that enforces strict isolation between public and private contexts. Having served paying users since April 4, 2025, the platform demonstrates that these five capabilities can coexist in a single maintainable production system — not merely as research prototypes — and remains freely accessible at groupgpt.tech.