Pipecat + MemorySync

Overview

Install pipecat-memorysync and place one FrameProcessor between your context aggregator and LLM service: every LLMContextFrame is enriched with relevant memories under a hard budget (default 1.2s) and its new user turns are sent to fact extraction, then the frame is pushed on time, enriched or not — the current utterance is in flight before the LLM replies. Capture is delta-only with deterministic idempotency seeds: growing a 50-message context does not re-send 50 messages per turn (a real flaw in the in-tree alternative, which re-sends the entire context every frame), and assistant messages are never sent. On EndFrame, queued sends get a bounded window to land, so the caller’s last words are not lost. Built on Pipecat 1.x’s current LLMContextFrame API and tested through pipecat.tests.utils.run_test — the framework’s own harness.

Setup status and requirements

Supported version

pipecat-memorysync 1.2.1 (PyPI)

Last setup review

2026-10-01

Permissions

A MemorySync API key with read and write scopes.

Limits

Requires pipecat-ai 1.0+ (Python 3.10+). Place after context_aggregator.user() and before the LLM service — enrichment reaches the model only from that position.

Capabilities

• Hard recall budget (default 1.2s) — the frame is never stalled
• Delta-only capture: new user messages only, never the entire context
• Deterministic idempotency seeds — a retried turn is extracted once
• Injected memory block is capture-excluded, by test
• Graceful EndFrame flush: the caller’s last words are not lost
• Built on the current LLMContextFrame API (Pipecat 1.x)
• Frame tests run through Pipecat’s official run_test harness
• Facts carry the pipecat:: session scope — same shared memories

Quick Start

from pipecat.pipeline.pipeline import Pipeline
from pipecat_memorysync import MemorySyncMemoryService

memory = MemorySyncMemoryService(api_key="ms_...",
                                 user_id="caller-42",
                                 session_id="call-123")

pipeline = Pipeline([
    transport.input(), stt,
    context_aggregator.user(),
    memory,               # between aggregator and LLM
    llm, tts,
    transport.output(),
    context_aggregator.assistant(),
])

Use Cases

Explore More