livekit-memorysync 1.0.0__tar.gz

livekit_memorysync-1.0.0/.gitignore

@@ -0,0 +1,5 @@
+venv/
+__pycache__/
+.pytest_cache/
+dist/
+*.egg-info/

livekit_memorysync-1.0.0/PKG-INFO

@@ -0,0 +1,141 @@
+Metadata-Version: 2.5
+Name: livekit-memorysync
+Version: 1.0.0
+Summary: MemorySync for LiveKit Agents: voice agents that remember callers — budgeted recall injection (never a stalled reply), both-role capture with idempotency seeds, and background prefetch.
+Project-URL: Homepage, https://docs.memorysync.io/guides/livekit
+Project-URL: Documentation, https://docs.memorysync.io/guides/livekit
+Project-URL: Repository, https://github.com/Rafay121/memorysync-plugins
+Author-email: MemorySync <support@memorysync.io>
+License-Expression: MIT
+Keywords: agents,livekit,long-term-memory,memory,memorysync,voice
+Classifier: Development Status :: 5 - Production/Stable
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Communications :: Conferencing
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: httpx<1,>=0.25
+Requires-Dist: livekit-agents<2,>=1.0.0
+Description-Content-Type: text/markdown
+
+# livekit-memorysync
+
+[MemorySync](https://memorysync.io) for [LiveKit Agents](https://docs.livekit.io/agents/) —
+voice agents that remember callers across calls, without ever stalling a reply.
+
+```bash
+pip install livekit-memorysync
+```
+
+## Why this exists
+
+Voice is the one surface where memory latency is *audible*. A text chatbot can
+spend two seconds fetching context; a voice agent that does so sounds broken.
+This package is built around that constraint:
+
+- **Budgeted recall.** Memory context is injected in `on_user_turn_completed`
+  under a hard timeout (default **1.2 s**). If MemorySync doesn't answer in
+  time, the reply proceeds *without* memories — never late.
+- **Background prefetch.** After each turn, the next recall is warmed in the
+  background, so the common case is an instant cache hit, not a network call.
+- **Both-role capture.** User *and* assistant turns are persisted (with
+  interruption metadata) via `conversation_item_added` — competitors that only
+  store user turns lose half the conversation.
+- **Delta-only, idempotent writes.** Every stored turn carries a deterministic
+  seed, so retries and reconnects never duplicate memories.
+- **Failure-proof.** Memory outages, quota limits, and dead networks degrade to
+  "no memories this turn". The call itself is never affected.
+
+## Quick start (composition — recommended)
+
+Keep your own `Agent` subclass; attach memory to it:
+
+```python
+from livekit.agents import Agent, AgentSession
+from livekit_memorysync import MemorySyncMemory
+
+memory = MemorySyncMemory(
+    api_key="ms_...",              # or MEMORYSYNC_API_KEY env var
+    user_id="caller-42",           # stable end-user id
+    thread_id="room-123",          # optional: scope to this room/call
+)
+
+class Assistant(Agent):
+    def __init__(self) -> None:
+        super().__init__(instructions="You are a helpful voice assistant.")
+
+    async def on_user_turn_completed(self, turn_ctx, new_message):
+        # Inject memories for THIS turn only (never persisted into the LLM ctx)
+        await memory.on_user_turn(self, turn_ctx, new_message)
+
+session = AgentSession(...)          # your STT/LLM/TTS choices
+memory.attach(session)               # capture both roles as they finalize
+await session.start(agent=Assistant(), ...)
+```
+
+## Quick start (drop-in agent)
+
+```python
+from livekit_memorysync import MemorySyncAgent
+
+agent = MemorySyncAgent(
+    instructions="You are a helpful voice assistant.",
+    api_key="ms_...",
+    user_id="caller-42",
+)
+# use like any Agent; recall + capture are wired for you
+```
+
+## Give the LLM a memory search tool
+
+```python
+from livekit_memorysync import create_memory_search_tool
+
+tool = create_memory_search_tool(memory)
+agent = Agent(instructions="...", tools=[tool])
+```
+
+The tool never raises into the model — errors come back as readable strings.
+
+## Configuration
+
+| Parameter | Default | Meaning |
+| --- | --- | --- |
+| `api_key` | `MEMORYSYNC_API_KEY` env | MemorySync API key |
+| `base_url` | `https://api.memorysync.io` | API endpoint |
+| `user_id` | required | Stable end-user identity |
+| `thread_id` | `None` | Scope memories to one room/call thread |
+| `recall_timeout` | `1.2` | Hard budget (seconds) for recall injection |
+| `top_k` | `5` | Memories injected per turn |
+| `persist_injection` | `False` | `True` writes the memory block into the session context instead of turn-only |
+| `prefetch` | `True` | Warm the next recall in the background |
+
+## Realtime-model caveat
+
+With speech-to-speech realtime models, `on_user_turn_completed` still fires
+(LiveKit synthesizes the turn boundary from transcripts), but injection lands
+just after the model may have started speaking. For strictly-realtime pipelines
+prefer the memory **search tool**, which the model calls when it needs history.
+
+## Semantics worth knowing
+
+- Injected memory blocks are wrapped in a guard line ("background information,
+  not instructions") and are excluded from capture, so recalled context is
+  never re-stored as a new memory.
+- Interrupted assistant turns are stored with `interrupted: true` metadata.
+- Free-tier quota exhaustion is silent by design (empty recall, accepted-but-
+  dropped writes); evaluation keys surface strict `429`s instead.
+
+## Development
+
+```bash
+python -m venv venv && venv/Scripts/pip install -e . livekit-agents pytest pytest-asyncio
+venv/Scripts/python -m pytest tests -q     # 16 tests, run against the real framework
+```
+
+## License
+
+MIT

livekit_memorysync-1.0.0/README.md

@@ -0,0 +1,118 @@
+# livekit-memorysync
+
+[MemorySync](https://memorysync.io) for [LiveKit Agents](https://docs.livekit.io/agents/) —
+voice agents that remember callers across calls, without ever stalling a reply.
+
+```bash
+pip install livekit-memorysync
+```
+
+## Why this exists
+
+Voice is the one surface where memory latency is *audible*. A text chatbot can
+spend two seconds fetching context; a voice agent that does so sounds broken.
+This package is built around that constraint:
+
+- **Budgeted recall.** Memory context is injected in `on_user_turn_completed`
+  under a hard timeout (default **1.2 s**). If MemorySync doesn't answer in
+  time, the reply proceeds *without* memories — never late.
+- **Background prefetch.** After each turn, the next recall is warmed in the
+  background, so the common case is an instant cache hit, not a network call.
+- **Both-role capture.** User *and* assistant turns are persisted (with
+  interruption metadata) via `conversation_item_added` — competitors that only
+  store user turns lose half the conversation.
+- **Delta-only, idempotent writes.** Every stored turn carries a deterministic
+  seed, so retries and reconnects never duplicate memories.
+- **Failure-proof.** Memory outages, quota limits, and dead networks degrade to
+  "no memories this turn". The call itself is never affected.
+
+## Quick start (composition — recommended)
+
+Keep your own `Agent` subclass; attach memory to it:
+
+```python
+from livekit.agents import Agent, AgentSession
+from livekit_memorysync import MemorySyncMemory
+
+memory = MemorySyncMemory(
+    api_key="ms_...",              # or MEMORYSYNC_API_KEY env var
+    user_id="caller-42",           # stable end-user id
+    thread_id="room-123",          # optional: scope to this room/call
+)
+
+class Assistant(Agent):
+    def __init__(self) -> None:
+        super().__init__(instructions="You are a helpful voice assistant.")
+
+    async def on_user_turn_completed(self, turn_ctx, new_message):
+        # Inject memories for THIS turn only (never persisted into the LLM ctx)
+        await memory.on_user_turn(self, turn_ctx, new_message)
+
+session = AgentSession(...)          # your STT/LLM/TTS choices
+memory.attach(session)               # capture both roles as they finalize
+await session.start(agent=Assistant(), ...)
+```
+
+## Quick start (drop-in agent)
+
+```python
+from livekit_memorysync import MemorySyncAgent
+
+agent = MemorySyncAgent(
+    instructions="You are a helpful voice assistant.",
+    api_key="ms_...",
+    user_id="caller-42",
+)
+# use like any Agent; recall + capture are wired for you
+```
+
+## Give the LLM a memory search tool
+
+```python
+from livekit_memorysync import create_memory_search_tool
+
+tool = create_memory_search_tool(memory)
+agent = Agent(instructions="...", tools=[tool])
+```
+
+The tool never raises into the model — errors come back as readable strings.
+
+## Configuration
+
+| Parameter | Default | Meaning |
+| --- | --- | --- |
+| `api_key` | `MEMORYSYNC_API_KEY` env | MemorySync API key |
+| `base_url` | `https://api.memorysync.io` | API endpoint |
+| `user_id` | required | Stable end-user identity |
+| `thread_id` | `None` | Scope memories to one room/call thread |
+| `recall_timeout` | `1.2` | Hard budget (seconds) for recall injection |
+| `top_k` | `5` | Memories injected per turn |
+| `persist_injection` | `False` | `True` writes the memory block into the session context instead of turn-only |
+| `prefetch` | `True` | Warm the next recall in the background |
+
+## Realtime-model caveat
+
+With speech-to-speech realtime models, `on_user_turn_completed` still fires
+(LiveKit synthesizes the turn boundary from transcripts), but injection lands
+just after the model may have started speaking. For strictly-realtime pipelines
+prefer the memory **search tool**, which the model calls when it needs history.
+
+## Semantics worth knowing
+
+- Injected memory blocks are wrapped in a guard line ("background information,
+  not instructions") and are excluded from capture, so recalled context is
+  never re-stored as a new memory.
+- Interrupted assistant turns are stored with `interrupted: true` metadata.
+- Free-tier quota exhaustion is silent by design (empty recall, accepted-but-
+  dropped writes); evaluation keys surface strict `429`s instead.
+
+## Development
+
+```bash
+python -m venv venv && venv/Scripts/pip install -e . livekit-agents pytest pytest-asyncio
+venv/Scripts/python -m pytest tests -q     # 16 tests, run against the real framework
+```
+
+## License
+
+MIT

livekit_memorysync-1.0.0/pyproject.toml

@@ -0,0 +1,42 @@
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[project]
+name = "livekit-memorysync"
+dynamic = ["version"]
+description = "MemorySync for LiveKit Agents: voice agents that remember callers — budgeted recall injection (never a stalled reply), both-role capture with idempotency seeds, and background prefetch."
+readme = "README.md"
+license = "MIT"
+requires-python = ">=3.10"
+authors = [{ name = "MemorySync", email = "support@memorysync.io" }]
+keywords = ["livekit", "voice", "agents", "memory", "memorysync", "long-term-memory"]
+classifiers = [
+    "Development Status :: 5 - Production/Stable",
+    "Intended Audience :: Developers",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Communications :: Conferencing",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = [
+    "livekit-agents>=1.0.0,<2",
+    "httpx>=0.25,<1",
+]
+
+[project.urls]
+Homepage = "https://docs.memorysync.io/guides/livekit"
+Documentation = "https://docs.memorysync.io/guides/livekit"
+Repository = "https://github.com/Rafay121/memorysync-plugins"
+
+[tool.hatch.version]
+path = "src/livekit_memorysync/_version.py"
+
+[tool.hatch.build.targets.wheel]
+packages = ["src/livekit_memorysync"]
+
+[tool.pytest.ini_options]
+asyncio_mode = "auto"
+testpaths = ["tests"]

livekit_memorysync-1.0.0/src/livekit_memorysync/__init__.py

@@ -0,0 +1,40 @@
+"""MemorySync for LiveKit Agents — voice agents that remember callers.
+
+Two ways in, one engine:
+
+- **Composition (recommended)** — keep YOUR Agent subclass; wire two lines::
+
+      memory = MemorySyncMemory(api_key=..., user_id="caller-123", thread_id=room_name)
+
+      class MyAgent(Agent):
+          async def on_user_turn_completed(self, turn_ctx, new_message):
+              await memory.on_user_turn(turn_ctx, new_message)
+
+      memory.attach(session)   # captures BOTH sides of the conversation
+
+- **Drop-in** — ``MemorySyncAgent`` subclasses ``Agent`` and wires it for you.
+
+The voice contract this package is built around: **recall runs under a hard
+time budget** (default 1.2 s) — a slow network yields a memoryless turn,
+never a stalled spoken reply — and **persistence never blocks anything**
+(fire-and-forget with content-hash idempotency seeds, both roles captured,
+interruptions marked). A background prefetch primes the next turn's recall,
+so steady-state injection costs ~0 ms.
+"""
+
+from ._api import AsyncV1Api, MemorySyncAPIError, fnv1a64
+from ._version import __version__
+from .memory import MemorySyncMemory, MemoryInjection
+from .agent import MemorySyncAgent
+from .tools import create_memory_search_tool
+
+__all__ = [
+    "MemorySyncMemory",
+    "MemorySyncAgent",
+    "MemoryInjection",
+    "create_memory_search_tool",
+    "AsyncV1Api",
+    "MemorySyncAPIError",
+    "fnv1a64",
+    "__version__",
+]

livekit_memorysync-1.0.0/src/livekit_memorysync/_api.py

@@ -0,0 +1,295 @@
+"""Async client for the MemorySync v1 data plane used by this integration.
+
+Conversation turns persist through the *episodic* ingestion path
+(``POST /v1/memory/add_turn``), which stores text verbatim — no fact
+extraction, no low-value-chatter gate, no rewriting. A voice transcript
+must round-trip byte-for-byte; a plane that second-guessed it would
+corrupt the caller's history.
+
+Everything here is async-native because LiveKit's agent runtime drives
+every call from the event loop — a blocking HTTP client inside a turn
+hook would stall the spoken reply.
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Any, Dict, List, Optional
+
+import httpx
+
+from ._version import __version__
+
+DEFAULT_BASE_URL = "https://api.memorysync.io"
+_USER_AGENT = f"livekit-memorysync/{__version__}"
+
+#: Namespace used when the key cannot list projects (see resolve_tenant_id).
+FALLBACK_TENANT = "default"
+
+
+class MemorySyncAPIError(Exception):
+    """A MemorySync call failed. Carries the status code and server detail."""
+
+    def __init__(self, message: str, *, status_code: Optional[int] = None) -> None:
+        super().__init__(message)
+        self.status_code = status_code
+
+
+def resolve_api_key(api_key: Optional[str]) -> str:
+    key = api_key or os.environ.get("MEMORYSYNC_API_KEY", "")
+    if not key or not key.strip():
+        raise ValueError(
+            "A MemorySync API key is required. Pass api_key=... or set the "
+            "MEMORYSYNC_API_KEY environment variable."
+        )
+    return key.strip()
+
+
+def resolve_base_url(base_url: Optional[str]) -> str:
+    url = base_url or os.environ.get("MEMORYSYNC_BASE_URL", "") or DEFAULT_BASE_URL
+    return url.rstrip("/")
+
+
+def fnv1a64(value: str) -> str:
+    """FNV-1a 64-bit over UTF-16 code units, as a fixed-width hex string.
+
+    Over UTF-16 code units — not code points, not UTF-8 bytes — so the
+    output matches the JavaScript adapters (`fnv1a64` in memorysync-ai-sdk
+    and memorysync-mastra hash over ``charCodeAt``) character for
+    character. Identical seeds across languages mean a turn persisted by a
+    Python surface and again by a JS surface converge on one stored row.
+    """
+    prime = 0x100000001B3
+    mask = 0xFFFFFFFFFFFFFFFF
+    h = 0xCBF29CE484222325
+    data = value.encode("utf-16-le")
+    for i in range(0, len(data), 2):
+        unit = data[i] | (data[i + 1] << 8)
+        h ^= unit
+        h = (h * prime) & mask
+    return format(h, "016x")
+
+
+class AsyncV1Api:
+    """Minimal asynchronous v1 client: add_turn, recall, query, list."""
+
+    def __init__(
+        self,
+        *,
+        api_key: str,
+        base_url: str,
+        project_id: Optional[str] = None,
+        timeout: float = 30.0,
+        transport: Optional[httpx.AsyncBaseTransport] = None,
+    ) -> None:
+        self._api_key = api_key
+        self._base_url = base_url.rstrip("/")
+        self._project_id = project_id
+        self._timeout = timeout
+        self._transport = transport
+        self._http = httpx.AsyncClient(timeout=timeout, transport=transport)
+        self._tenant_id: Optional[str] = None
+        self._tenant_is_fallback = False
+
+    async def aclose(self) -> None:
+        await self._http.aclose()
+
+    @property
+    def api_key(self) -> str:
+        return self._api_key
+
+    @property
+    def base_url(self) -> str:
+        return self._base_url
+
+    @property
+    def project_id(self) -> Optional[str]:
+        return self._project_id
+
+    @property
+    def timeout(self) -> float:
+        return self._timeout
+
+    @property
+    def transport(self) -> Optional[httpx.AsyncBaseTransport]:
+        return self._transport
+
+    # ── plumbing ─────────────────────────────────────────────────────
+
+    def _headers(self) -> Dict[str, str]:
+        h = {
+            "X-API-Key": self._api_key,
+            "Accept": "application/json",
+            "User-Agent": _USER_AGENT,
+        }
+        if self._project_id:
+            h["X-Project-ID"] = self._project_id
+        return h
+
+    async def _request(
+        self,
+        method: str,
+        path: str,
+        *,
+        json: Optional[Dict[str, Any]] = None,
+        params: Optional[Dict[str, Any]] = None,
+    ) -> Any:
+        url = f"{self._base_url}{path}"
+        try:
+            response = await self._http.request(
+                method, url, headers=self._headers(), json=json, params=params
+            )
+        except httpx.TimeoutException as e:
+            raise MemorySyncAPIError(f"Request timed out: {e}") from e
+        except httpx.HTTPError as e:
+            raise MemorySyncAPIError(f"Network error: {e}") from e
+
+        if response.status_code == 204:
+            return None
+        try:
+            body: Any = response.json()
+        except ValueError:
+            body = response.text or None
+        if response.status_code >= 400:
+            detail = body.get("detail") if isinstance(body, dict) else body
+            raise MemorySyncAPIError(
+                f"{method} {path} failed with HTTP {response.status_code}: {detail}",
+                status_code=response.status_code,
+            )
+        return body
+
+    # ── calls ────────────────────────────────────────────────────────
+
+    async def resolve_tenant_id(self) -> str:
+        """The tenant id, which the v1 routes need in path or body.
+
+        Derived from the project listing rather than asked for. Cached for
+        the lifetime of this client; one extra GET per process, not per
+        turn. Keys without the ``projects:read`` scope (evaluation keys)
+        fall back to the fixed namespace ``"default"`` — deterministic, so
+        every read and write through this client lands in one namespace.
+        Only a definite 401/403 triggers the fallback; a transient server
+        error re-raises rather than silently switching namespaces.
+        """
+        if self._tenant_id:
+            return self._tenant_id
+        try:
+            projects = await self._request("GET", "/org/projects")
+        except MemorySyncAPIError as exc:
+            if exc.status_code in (401, 403):
+                self._tenant_id = FALLBACK_TENANT
+                self._tenant_is_fallback = True
+                return self._tenant_id
+            raise
+        first = projects[0] if isinstance(projects, list) and projects else None
+        tenant = first.get("tenant_id") if isinstance(first, dict) else None
+        if not tenant:
+            raise MemorySyncAPIError(
+                "Could not determine the tenant for this API key. Pass "
+                "tenant_id=... explicitly, or verify the key with "
+                "`memorysync doctor`."
+            )
+        self._tenant_id = str(tenant)
+        return self._tenant_id
+
+    @property
+    def tenant_is_fallback(self) -> bool:
+        """True when the namespace came from the 401/403 fallback."""
+        return self._tenant_is_fallback
+
+    def set_tenant_id(self, tenant_id: str) -> None:
+        self._tenant_id = tenant_id
+
+    async def add_turn(
+        self,
+        *,
+        tenant_id: str,
+        user_id: str,
+        text: str,
+        speaker: Optional[str] = None,
+        occurred_at: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+        source: str = "livekit",
+        sync_embed: bool = False,
+    ) -> Dict[str, Any]:
+        """Store one item verbatim (episodic ingestion).
+
+        ``speaker`` + ``occurred_at`` participate in the server's
+        idempotency seed, so retrying an identical payload is recognised
+        (``already_exists: true``) instead of stored twice.
+        """
+        body: Dict[str, Any] = {
+            "tenant_id": tenant_id,
+            "user_id": user_id,
+            "source": source,
+            "text": text,
+            "sync_embed": sync_embed,
+        }
+        if speaker is not None:
+            body["speaker"] = speaker
+        if occurred_at is not None:
+            body["occurred_at"] = occurred_at
+        if metadata is not None:
+            body["metadata"] = metadata
+        return await self._request("POST", "/v1/memory/add_turn", json=body) or {}
+
+    async def recall(
+        self,
+        *,
+        tenant_id: str,
+        user_id: str,
+        prompt: str,
+        k: Optional[int] = None,
+        types: Optional[List[str]] = None,
+    ) -> Dict[str, Any]:
+        """Hierarchical recall: grouped, prompt-ready context block."""
+        body: Dict[str, Any] = {
+            "tenant_id": tenant_id,
+            "user_id": user_id,
+            "prompt": prompt,
+        }
+        if k is not None:
+            body["k"] = k
+        if types is not None:
+            body["types"] = types
+        return await self._request("POST", "/v1/memory/recall", json=body) or {}
+
+    async def query(
+        self,
+        *,
+        tenant_id: str,
+        user_id: str,
+        prompt: str,
+        k: Optional[int] = None,
+    ) -> Dict[str, Any]:
+        """Plain semantic search over the pair's memories (episodic included)."""
+        body: Dict[str, Any] = {
+            "tenant_id": tenant_id,
+            "user_id": user_id,
+            "prompt": prompt,
+        }
+        if k is not None:
+            body["k"] = k
+        return await self._request("POST", "/v1/memory/query", json=body) or {}
+
+    async def list_memories(
+        self,
+        *,
+        tenant_id: str,
+        user_id: str,
+        limit: int = 0,
+    ) -> List[Dict[str, Any]]:
+        """Every memory for the tenant/user pair, newest first.
+
+        ``limit=0`` means no limit — a transcript read must never be
+        silently truncated, so that is the default here.
+        """
+        from urllib.parse import quote
+
+        raw = await self._request(
+            "GET",
+            f"/v1/memory/{quote(tenant_id, safe='')}/{quote(user_id, safe='')}/list",
+            params={"limit": limit},
+        )
+        memories = raw.get("memories") if isinstance(raw, dict) else None
+        return list(memories) if isinstance(memories, list) else []

livekit_memorysync-1.0.0/src/livekit_memorysync/_version.py

@@ -0,0 +1,3 @@
+"""Version for livekit-memorysync. Single source of truth."""
+
+__version__ = "1.0.0"