sensorium-mcp

MCP server with 5-layer memory, voice analysis, 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

Quickstart

npx sensorium-mcp@latest

Or add to your VS Code mcp.json:

{
  "servers": {
    "sensorium-mcp": {
      "command": "npx",
      "args": ["sensorium-mcp@latest"],
      "env": {
        "TELEGRAM_TOKEN": "...",
        "TELEGRAM_CHAT_ID": "...",
        "OPENAI_API_KEY": "...",
        "WAIT_TIMEOUT_MINUTES": "30"
      }
    }
  }
}

Then tell your agent:

Start remote copilot session

Features

5-Layer Memory System

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

| Layer | What it stores | |-------|---------------| | Working Memory | Current session context — active goals, recent messages | | Episodic Memory | Auto-saved conversation episodes (every operator message) | | Semantic Memory | Extracted facts, preferences, patterns, entities, relationships | | Procedural Memory | Multi-step procedures and workflows | | Meta-Memory | Confidence scores, decay tracking, topic indexing |

Storage: SQLite at ~/.sensorium-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.

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 VS Code restarts
• Image, document, and video note support
• Voice messages with Whisper transcription
• Automatic Markdown → Telegram MarkdownV2 conversion

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

Dead Session Detection

Automatic alert when no tool calls arrive for 60 minutes. Single alert per downtime — no spam. Replaces annoying keep-alive pings.

Tools

| Tool | Description | |------|-------------| | start_session | Begin or resume a session with optional memory bootstrap | | remote_copilot_wait_for_instructions | Block until operator message, scheduled task, or timeout | | report_progress | Send Markdown progress update to operator | | send_file | Send file or image to operator | | send_voice | Text-to-speech voice message via OpenAI TTS | | schedule_wake_up | Schedule a one-shot or idle task | | memory_bootstrap | Load memory briefing into context | | memory_search | Search episodic/semantic memory by query | | memory_save | Save a fact, preference, pattern, entity, or relationship | | memory_save_procedure | Save a multi-step procedure | | memory_update | Update or supersede an existing note | | memory_consolidate | Run intelligent consolidation | | memory_status | Check memory health and statistics | | memory_forget | Delete a specific memory note |

Data Privacy

Memory Consolidation

The memory system periodically sends conversation excerpts to OpenAI's API for knowledge extraction and consolidation. This helps maintain useful context across sessions.

To disable this behavior, set the environment variable:

CONSOLIDATION_ENABLED=false

When disabled, the memory system will still store episodes locally but will not send them to OpenAI for consolidation.

Environment Variables

| Variable | Required | Default | Description | |----------|----------|---------|-------------| | TELEGRAM_TOKEN | Yes | — | Telegram Bot API token | | TELEGRAM_CHAT_ID | Yes | — | Forum supergroup chat ID | | OPENAI_API_KEY | No | — | For voice transcription (Whisper), TTS, and memory consolidation | | VOICE_ANALYSIS_URL | No | — | Voice emotion analysis microservice URL | | CONSOLIDATION_ENABLED | No | true | Set to false or 0 to disable sending episodes to OpenAI for consolidation | | CONSOLIDATION_MODEL | No | gpt-4o-mini | OpenAI model for memory consolidation | | MCP_HTTP_PORT | No | — | If set, starts HTTP/SSE transport on this port instead of stdio | | WAIT_TIMEOUT_MINUTES | No | 120 | Wait timeout in minutes |

Prerequisites

• Node.js 18+ (uses native fetch)
• A Telegram bot token
• A Telegram forum supergroup with the bot as admin (Manage Topics right)

Transport Modes

stdio (default) — standard MCP transport. Used with npx sensorium-mcp@latest.

HTTP/SSE — set MCP_HTTP_PORT to start an HTTP server. Useful for development (restart server without restarting VS Code) or remote connections:

{
  "servers": {
    "sensorium-mcp": {
      "type": "streamableHttp",
      "url": "http://localhost:3847/mcp"
    }
  }
}

Start the server separately:

MCP_HTTP_PORT=3847 TELEGRAM_TOKEN=... TELEGRAM_CHAT_ID=... node dist/index.js

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 and memory consolidation extracts durable knowledge from episodes.

Architecture

~/.sensorium-mcp/
  poller.lock                 ← PID + timestamp; first instance becomes the poller
  offset                      ← shared getUpdates offset
  memory.db                   ← SQLite: episodes, semantic notes, procedures, voice signatures
  threads/
    <threadId>.jsonl           ← messages for each topic thread
    general.jsonl              ← messages with no thread ID
~/.sensorium-mcp-sessions.json  ← name → threadId mapping

License

MIT