tibet-voice-cache-mcp

MCP server for persistent voice conversation memory. Plug into Claude Code, Cursor, Windsurf, or any MCP client.

pip install tibet-voice-cache-mcp

What it does

Gives any MCP-compatible AI client tools to store and recall voice conversation context. User and AI utterances are stored separately in RAM (or optionally on disk) and formatted as clean context summaries — no fake turns, no role confusion.

┌──────────────────────────────────────────────────────────────┐
│  MCP Client (Claude Code / Cursor / Windsurf / etc.)         │
│                                                              │
│  voice_cache_add(actor="user_1", text="...", role="user")    │
│  voice_cache_add(actor="user_1", text="...", role="ai")      │
│  voice_cache_turn(actor="user_1")                            │
│                                                              │
│  voice_cache_inject(actor="user_1",                          │
│    base_instruction="You are a voice assistant.")            │
│  → "You are a voice assistant.                               │
│                                                              │
│     === PRIOR CONTEXT ===                                    │
│     The user previously said:                                │
│       - What's the weather?                                  │
│     You previously responded:                                │
│       - Sunny and 22 degrees!                                │
│     === END CONTEXT ==="                                     │
└──────────────────────────────────────────────────────────────┘

Setup

Claude Code

// ~/.claude.json
{
  "mcpServers": {
    "voice-cache": {
      "command": "tibet-voice-cache-mcp"
    }
  }
}

With disk persistence

{
  "mcpServers": {
    "voice-cache": {
      "command": "tibet-voice-cache-mcp",
      "env": {
        "VOICE_CACHE_DIR": "/path/to/cache"
      }
    }
  }
}

Cursor / Windsurf

Same pattern — add tibet-voice-cache-mcp as an MCP server command.

Tools

Tool	Description
voice_cache_status	List all active caches with stats
voice_cache_open	Open/create cache for an actor
voice_cache_add	Record user or AI utterance
voice_cache_turn	Mark turn boundary
voice_cache_context	Get formatted context summary
voice_cache_inject	Inject context into system instruction
voice_cache_session	Bulk import session transcripts
voice_cache_history	View cached utterances
voice_cache_clear	Clear cache for an actor
voice_cache_configure	Change summary style / language

Quick workflow

# During voice session
voice_cache_open(actor="user_123")
voice_cache_add(actor="user_123", text="What's the weather?", role="user")
voice_cache_add(actor="user_123", text="Sunny and warm!", role="ai")
voice_cache_turn(actor="user_123")

# Next session — inject memory
voice_cache_inject(
    actor="user_123",
    base_instruction="You are a friendly weather assistant."
)

Summary styles

Configure how context is formatted:

voice_cache_configure(actor="user_123", summary_style="compact")

Style	Format
labeled	Sectioned with headers (default)
compact	Minimal tokens, single-line
narrative	Natural language, conversational
chronological	Numbered turn pairs

Multi-language

voice_cache_configure(actor="user_123", language="nl")

Built-in: English (en), Dutch (nl).

Environment variables

Variable	Default	Description
VOICE_CACHE_DIR	(none — RAM only)	Directory for JSON persistence
VOICE_CACHE_MAX_TURNS	50	Max utterances per side before trimming
VOICE_CACHE_STYLE	labeled	Default summary style

Resources

The server also exposes MCP resources:

• voice-cache://actors — List all actors with open caches
• voice-cache://actor/{name} — Full cache content for an actor

Part of the TIBET ecosystem

Package	Description
tibet-voice-cache	Core library — voice conversation memory
tibet-voice-cache-mcp	This package — MCP server wrapper

License

MIT — plug it in, give your voice AI a memory.