SaveMemoryFrame only after the user chooses text
to preserve, or a RecallMemoryFrame with a query and an observation ID from
a prior search or save. Ordinary audio and transcription frames pass through
without being stored, even if the transcript says “remember this.”
The example uses Pipecat’s custom frame processor contract
and the official Python MCP client.
It calls the existing local worker through Claude-Mem’s MCP server. There is
no new memory store, background capture, transcript replay or audio retention.
Requirements
The example is verified on macOS arm64 with Python 3.12.4, Pipecat 1.16.0, MCP SDK 2.3.0, Node.js 25.1.0 and Claude-Mem 15.18.0. It needs Python 3.11 or newer. Other operating systems have not been verified here. The runnable demonstration uses a real Pipecat pipeline with application-created frames. It does not open a microphone or contact an STT, LLM or TTS provider, so it needs no provider account, API key or CMEM Pro subscription. Audio transcription and speech playback in a full voice bot require your chosen Pipecat services; those integrations are not verified by this demonstration.Connect and verify
First install Claude-Mem, start the local worker and check its status. Copyexamples/pipecat from the repository, then install its
pinned Python dependencies in a virtual environment:
plugin/scripts/mcp-server.cjs. If your worker has a custom
CLAUDE_MEM_DATA_DIR, host or port, pass those same settings in this shell.
The example disables worker autostart and passes the existing environment to
the MCP subprocess; it does not edit any application configuration.
Save only a harmless, selected portion of a transcript:
success: true only after the worker returns a save receipt.
Use that receipt’s ID in a later recall, replacing 123 below:
Add the stage to a voice application
ProjectMemory emits a MemoryResultFrame with text and success.
Your application can display it or route the text to its speech service.
Treat recalled text as reference material, not authorization to execute tools.
Do not translate arbitrary model output or transcript text into save frames:
the host application owns selection and consent.
Privacy, failures and disconnect
Only selected save text, its title, the chosen project and deduplication metadata are stored. Nothing records audio. No full transcript, user identity or conversation history is sent automatically. The existing worker’s indexing, sync and export settings still apply to explicitly saved notes. Connection and tool requests have a ten-second deadline; the MCP SDK separately bounds subprocess shutdown. There are no automatic retries or deferred writes. A failed or ambiguous save producessuccess: false; check the connection before explicitly trying again. The
same fingerprint makes an exact retry idempotent if a receipt was lost.
Failures do not prevent later audio or transcript frames from passing through.
Set enabled=False or remove the stage to disconnect. Each explicit operation
closes its MCP connection. No hook or configuration entry needs uninstalling;
disconnecting leaves previously saved notes intact.
Tests
verify_worker.py uses the real worker/MCP
bundle, verifies an exact replay and project isolation, and reports IDs and
failures. Do not point it at a personal project unless you want its test notes
to remain in that worker.