sensorium-mcp

MCP server with multi-layer memory, voice analysis, multi-thread orchestration, and Telegram bridge for AI assistants.

Why?

AI assistants forget everything between sessions. Every restart is a blank slate — no memory of your preferences, past decisions, or ongoing projects. Voice messages arrive as opaque audio blobs. And there's no way to talk to your agent when it's running headless in CI or on a remote machine.

sensorium-mcp fixes all three problems:

• Persistent memory that survives across sessions, automatically capturing episodes and consolidating knowledge

• Voice understanding with transcription and real-time emotion analysis

• Remote control via Telegram — give instructions, send files, receive progress updates from anywhere

• Multi-thread orchestration — spawn worker, branch, and daily threads with shared or isolated memory

Related MCP server: Voice-AGI MCP Server

Getting Started

Prerequisites

• Windows 10/11 (supervisor is Windows-only; the MCP server itself is cross-platform)

• Node.js 18+

• A Telegram bot token

• A Telegram forum supergroup with the bot as admin (Manage Topics right)

1. Create a Telegram bot and forum group

1. Message @BotFather on Telegram and create a new bot — copy the token

2. Create a Telegram group, convert to supergroup, and enable Topics in settings

3. Add your bot as admin with Manage Topics permission

4. Get the chat ID (use @userinfobot or the Telegram API)

2. Create a .env file

Copy .env.example to .env in the directory where you'll run the installer:

cp .env.example .env

Fill in at minimum TELEGRAM_TOKEN and TELEGRAM_CHAT_ID. Add OPENAI_API_KEY for voice and memory consolidation features. Set MCP_HTTP_PORT (e.g. 3847) to enable multi-thread agent spawning.

If you have SecureVault installed, the installer uses it automatically instead of .env.

3. Install and run

powershell -ExecutionPolicy Bypass -File .\Install-Sensorium.ps1

This downloads the supervisor binary, installs a startup launcher, loads secrets (from SecureVault or .env), and starts the supervisor in the background. The supervisor manages the MCP server lifecycle — spawning, health checks, auto-restart on crash, and coordinating updates.

Configuration is stored in ~/.remote-copilot-mcp/install.config.json.

4. Use it

The supervisor starts the MCP server on the configured HTTP port. When the server spawns agent threads (Claude, Copilot, Codex), it automatically generates per-thread MCP configs and injects them — no manual mcp.json setup needed.

Tell your agent:

Start remote copilot session

Running without the supervisor

For development or quick testing, run the MCP server directly:

# HTTP transport (recommended — required for multi-thread spawning)
MCP_HTTP_PORT=3847 TELEGRAM_TOKEN=... TELEGRAM_CHAT_ID=... npx sensorium-mcp@latest

# stdio transport (simplest, for single-session use)
npx sensorium-mcp@latest

Features

Multi-Layer Memory System

Every operator message is automatically captured. Knowledge is extracted and consolidated during idle time using a configurable LLM (default: gpt-4o).

Storage: SQLite at ~/.remote-copilot-mcp/memory.db. No external database required.

Auto-bootstrap — session start auto-injects a memory briefing so the agent immediately knows who you are and what you've been working on.

Auto-ingest — every operator message is saved as an episode automatically.

Intelligent consolidation — a configurable LLM analyzes accumulated episodes and extracts durable knowledge (facts, preferences, patterns) during idle periods. Includes deduplication, quality scoring, and causal linking.

Remote Control via Telegram

Operate your AI assistant from anywhere through a Telegram forum supergroup.

• Concurrent sessions with a shared file-based dispatcher (no 409 conflicts)

• Named session persistence across restarts

• Image, document, and video note support

• Voice messages with Whisper transcription

• Automatic Markdown to Telegram MarkdownV2 conversion

Multi-Thread Orchestration

Spawn and manage multiple agent threads from a single session.

Threads can communicate via send_message_to_thread and coordinate work across multiple agents.

Voice Analysis

Real-time voice emotion analysis via an optional microservice (see voice-analysis/).

• Detects emotions, gender, arousal/dominance/valence

• Video note (circle video) support with audio extraction

• Deployable via Docker

Scheduler

Schedule tasks that fire during wait_for_instructions.

• One-shot: runAt — trigger at a specific time

• Cron: cron — recurring schedule using 5-field cron expressions

• Idle-triggered: afterIdleMinutes — trigger after N minutes of inactivity

Skills System

Customizable prompt templates that agents can discover and load on demand.

• search_skills — find relevant skills by keyword

• get_skill — load a specific skill template

• Templates stored in ~/.remote-copilot-mcp/templates/ with {{VARIABLE}} bindings

Tools

Environment Variables

Data Privacy

The memory system periodically sends conversation excerpts to OpenAI's API for knowledge extraction and consolidation. To disable:

CONSOLIDATION_ENABLED=false

When disabled, episodes are still stored locally but not sent to OpenAI.

How It Works

1. start_session creates a Telegram topic (or resumes one by name). Memory bootstrap auto-loads your context.

2. A shared dispatcher runs a single getUpdates poller (elected via lock file). Messages are written to per-thread JSONL files — each MCP instance reads its own.

3. Incoming messages (text, photo, document, voice, video note) are processed, transcribed, and delivered as MCP content blocks. Every operator message is auto-saved as an episode.

4. The agent works, calls report_progress / send_file / send_voice, and loops back to wait_for_instructions.

5. During idle periods, the scheduler fires pending tasks, memory consolidation extracts durable knowledge from episodes, and reflection/narrative pipelines generate temporal context.

Architecture

~/.remote-copilot-mcp/
  memory.db                   <- SQLite: episodes, semantic notes, voice signatures, thread registry
  settings.json               <- Per-thread agent types, keep-alive, conversation modes
  install.config.json          <- Installer configuration (SecureVault profile, update mode)
  poller.lock                 <- PID + timestamp; first instance becomes the poller
  offset                      <- Shared getUpdates offset
  server.pid                  <- Authoritative PID for supervisor and self-update
  templates/                  <- Skill templates (*.default.md)
  threads/
    <threadId>.jsonl           <- Messages for each topic thread
  bin/
    sensorium-supervisor.exe   <- Supervisor binary (Windows)
  logs/
    threads/                   <- Per-thread agent logs

License

MIT