memorysync-pipecat

MemorySync for Pipecat — long-term memory for voice pipelines that never stalls a reply and never re-sends a turn it already sent. PyPI package: pipecat-memorysync.

Built and maintained by the MemorySync team — MemorySync is our product, and this integration is actively maintained alongside it.

Tested with Pipecat v1.8.1 (pipecat-ai>=1.0.0,<2).

pip install pipecat-memorysync
# or
uv add pipecat-memorysync

Where it sits

MemorySyncMemoryService is a FrameProcessor. Place it between your context aggregator and your LLM service:

transport.input() → stt → context_aggregator.user()
    → MemorySyncMemoryService        ← enriches + captures here
    → llm → tts → transport.output() → context_aggregator.assistant()

from pipecat_memorysync import MemorySyncMemoryService

memory = MemorySyncMemoryService(
    api_key="ms_...",                 # or MEMORYSYNC_API_KEY env var
    user_id="caller-42",              # stable end-user id
    session_id="call-123",            # optional: scope to this call
)

pipeline = Pipeline([
    transport.input(),
    stt,
    context_aggregator.user(),
    memory,
    llm,
    tts,
    transport.output(),
    context_aggregator.assistant(),
])

Every LLMContextFrame that flows through is enriched with relevant memories (as a system message) and its new user turns are sent to fact extraction — then pushed on, enriched or not, on time.

Design guarantees

• Budgeted recall. Enrichment runs under a hard timeout (default 1.2 s). A slow or dead memory backend means an unenriched frame, never a stalled voice reply.
• Immediate, loss-proof capture. Every new user message is sent the moment its frame passes — the current utterance is in flight BEFORE the LLM replies, so a disconnect can never lose it. Filler filtering and fact extraction happen server-side: only the durable facts in what the caller said are stored as memories, not the turn text.
• Delta-only capture. Only user messages not seen before are sent, tracked by deterministic idempotency seeds; a retried turn is recognised server-side and not extracted twice. Growing a 50-message context does not re-send 50 messages per turn.
• Injection exclusion. The memory block this service adds is never sent back as a new turn.
• Graceful end, salvaged abort. On EndFrame, queued sends get a bounded window (3 s) to land before the pipeline stops — the caller's last words are not lost. On CancelFrame, the frame is pushed first and sends get a brief salvage window.
• Failure-proof. HTTP errors, quota limits, and timeouts all degrade to "no memories this turn". Nothing propagates into the pipeline.

Running the example

A single-file voice agent that remembers callers across calls lives in examples/foundational.py:

uv add pipecat-memorysync "pipecat-ai[deepgram,cartesia,openai,silero,runner,webrtc]"

export MEMORYSYNC_API_KEY=ms_...   # https://app.memorysync.io
export DEEPGRAM_API_KEY=...
export CARTESIA_API_KEY=...
export OPENAI_API_KEY=...

python examples/foundational.py

Open http://localhost:7860/client, tell the bot your name and a preference, hang up, and connect again — it remembers.

Configuration (InputParams)

from pipecat_memorysync import MemorySyncMemoryService

memory = MemorySyncMemoryService(
    api_key="ms_...",
    user_id="caller-42",
    params=MemorySyncMemoryService.InputParams(
        top_k=6,                  # memories injected per turn
        recall_timeout=1.2,       # hard budget, seconds
        add_as_system_message=True,
        position=1,               # index where the memory block is inserted
        min_prompt_chars=8,       # skip enrichment for shorter user prompts
    ),
)

Param	Default	Meaning
top_k	6	Memories injected per turn
recall_timeout	1.2	Hard recall budget in seconds
system_prompt	(guarded header)	Prefix line for the injected block; also the capture-exclusion marker
add_as_system_message	True	Inject as system (else as a user message)
position	1	Index in the message list where the block is inserted (0 = first; the default lands right after your system prompt)
min_prompt_chars	8	Skip recall for trivial prompts

Semantics worth knowing

• Only the caller's (user) turns are sent, as plain text with role: "user". The server extracts the durable facts they contain and stores only those, tagged with session_id: "pipecat::<session>". Assistant replies are not sent — they are not stored as memories — and filler such as "ok" or "thanks" stores nothing.
• Idempotency seeds let the server recognise retries and reconnects: the same turn is not extracted twice.
• Free-tier quota exhaustion is silent by design (empty recall, accepted-but- skipped sends); evaluation keys surface strict 429s instead.
• The service is reusable across pipeline runs; call await memory.aclose() from application shutdown if you want an explicit flush + client close.

Development

python -m venv venv && venv/Scripts/pip install -e . pipecat-ai pytest pytest-asyncio
venv/Scripts/python -m pytest tests -q    # 19 tests, run via pipecat's official test harness

Docs

Full guide: docs.memorysync.io/guides/pipecat

License

MIT