feat: Add AgentCore Memory Checkpointer

libs/langgraph-checkpoint-aws/langgraph_checkpoint_aws/__init__.py

@@ -2,5 +2,15 @@
 LangGraph Checkpoint AWS - A LangChain checkpointer implementation using Bedrock Session Management Service.
 """
 
+from langgraph_checkpoint_aws.agentcore.saver import (
+    AgentCoreMemorySaver,
+)
+
 __version__ = "0.1.2"
 SDK_USER_AGENT = f"LangGraphCheckpointAWS#{__version__}"
+
+# Expose the saver class at the package level
+__all__ = [
+    "AgentCoreMemorySaver",
+    "SDK_USER_AGENT",
+]

libs/langgraph-checkpoint-aws/langgraph_checkpoint_aws/agentcore/__init__.py

@@ -0,0 +1,3 @@
+from langgraph_checkpoint_aws.agentcore.saver import AgentCoreMemorySaver
+
+__all__ = ["AgentCoreMemorySaver"]

libs/langgraph-checkpoint-aws/langgraph_checkpoint_aws/agentcore/constants.py

@@ -0,0 +1,29 @@
+"""
+Constants and exceptions for AgentCore Memory Checkpoint Saver.
+"""
+
+EMPTY_CHANNEL_VALUE = "_empty"
+
+
+class AgentCoreMemoryError(Exception):
+    """Base exception for AgentCore Memory errors."""
+
+    pass
+
+
+class EventDecodingError(AgentCoreMemoryError):
+    """Raised when event decoding fails."""
+
+    pass
+
+
+class InvalidConfigError(AgentCoreMemoryError):
+    """Raised when configuration is invalid."""
+
+    pass
+
+
+class EventNotFoundError(AgentCoreMemoryError):
+    """Raised when expected event is not found."""
+
+    pass

libs/langgraph-checkpoint-aws/langgraph_checkpoint_aws/agentcore/helpers.py

@@ -0,0 +1,313 @@
+"""
+Helper classes for AgentCore Memory Checkpoint Saver.
+"""
+
+from __future__ import annotations
+
+import base64
+import datetime
+import json
+import logging
+from collections import defaultdict
+from typing import Any, Dict, List, Union
+
+import boto3
+from langgraph.checkpoint.base import CheckpointTuple, SerializerProtocol
+
+from langgraph_checkpoint_aws.agentcore.constants import (
+    EMPTY_CHANNEL_VALUE,
+    EventDecodingError,
+)
+from langgraph_checkpoint_aws.agentcore.models import (
+    ChannelDataEvent,
+    CheckpointerConfig,
+    CheckpointEvent,
+    WriteItem,
+    WritesEvent,
+)
+
+logger = logging.getLogger(__name__)
+
+# Union type for all events
+EventType = Union[CheckpointEvent, ChannelDataEvent, WritesEvent]
+
+
+class EventSerializer:
+    """Handles serialization and deserialization of events to store in AgentCore Memory."""
+
+    def __init__(self, serde: SerializerProtocol):
+        self.serde = serde
+
+    def serialize_value(self, value: Any) -> Dict[str, Any]:
+        """Serialize a value using the serde protocol."""
+        type_tag, binary_data = self.serde.dumps_typed(value)
+        return {"type": type_tag, "data": base64.b64encode(binary_data).decode("utf-8")}
+
+    def deserialize_value(self, serialized: Dict[str, Any]) -> Any:
+        """Deserialize a value using the serde protocol."""
+        try:
+            type_tag = serialized["type"]
+            binary_data = base64.b64decode(serialized["data"])
+            return self.serde.loads_typed((type_tag, binary_data))
+        except Exception as e:
+            raise EventDecodingError(f"Failed to deserialize value: {e}")
+
+    def serialize_event(self, event: EventType) -> str:
+        """Serialize an event to JSON string."""
+
+        # Create a custom serializer for Pydantic models
+        def custom_serializer(obj):
+            if hasattr(obj, "model_dump"):
+                return obj.model_dump()
+            raise TypeError(f"Object of type {type(obj)} is not JSON serializable")
+
+        # Get the base dictionary
+        event_dict = event.model_dump(exclude_none=True)
+
+        # Handle special serialization for specific fields
+        if isinstance(event, CheckpointEvent):
+            event_dict["checkpoint_data"] = self.serialize_value(event.checkpoint_data)
+            event_dict["metadata"] = self.serialize_value(event.metadata)
+
+        elif isinstance(event, ChannelDataEvent):
+            if event.value != EMPTY_CHANNEL_VALUE:
+                event_dict["value"] = self.serialize_value(event.value)
+
+        elif isinstance(event, WritesEvent):
+            # The writes field is already properly serialized by model_dump()
+            # We just need to serialize the value field in each write
+            for write in event_dict["writes"]:
+                val = write.get("value", EMPTY_CHANNEL_VALUE)
+                write["value"] = self.serialize_value(val)
+
+        return json.dumps(event_dict, default=custom_serializer)
+
+    def deserialize_event(self, data: str) -> EventType:
+        """Deserialize JSON string to event."""
+        try:
+            event_dict = json.loads(data)
+            event_type = event_dict.get("event_type")
+
+            if event_type == "checkpoint":
+                # Deserialize checkpoint data and metadata
+                event_dict["checkpoint_data"] = self.deserialize_value(
+                    event_dict["checkpoint_data"]
+                )
+                event_dict["metadata"] = self.deserialize_value(event_dict["metadata"])
+                return CheckpointEvent(**event_dict)
+
+            elif event_type == "channel_data":
+                # Deserialize channel value if not empty
+                if "value" in event_dict and isinstance(event_dict["value"], dict):
+                    event_dict["value"] = self.deserialize_value(event_dict["value"])
+                return ChannelDataEvent(**event_dict)
+
+            elif event_type == "writes":
+                # Deserialize write values
+                for write in event_dict["writes"]:
+                    if isinstance(write["value"], dict):
+                        write["value"] = self.deserialize_value(write["value"])
+                return WritesEvent(**event_dict)
+
+            else:
+                raise EventDecodingError(f"Unknown event type: {event_type}")
+
+        except json.JSONDecodeError as e:
+            raise EventDecodingError(f"Failed to parse JSON: {e}")
+        except Exception as e:
+            raise EventDecodingError(f"Failed to deserialize event: {e}")
+
+
+class AgentCoreEventClient:
+    """Handles low-level event storage and retrieval from AgentCore Memory for checkpoints."""
+
+    def __init__(self, memory_id: str, serializer: EventSerializer, **boto3_kwargs):
+        self.memory_id = memory_id
+        self.serializer = serializer
+        self.client = boto3.client("bedrock-agentcore", **boto3_kwargs)
+
+    def store_event(self, event: EventType, session_id: str, actor_id: str) -> None:
+        """Store an event in AgentCore Memory."""
+        serialized = self.serializer.serialize_event(event)
+
+        self.client.create_event(
+            memoryId=self.memory_id,
+            actorId=actor_id,
+            sessionId=session_id,
+            eventTimestamp=datetime.datetime.now(datetime.timezone.utc),
+            payload=[{"blob": serialized}],
+        )
+
+    def store_events_batch(
+        self, events: List[EventType], session_id: str, actor_id: str
+    ) -> None:
+        """Store multiple events in a single API call to AgentCore Memory."""
+        # Serialize all events into payload blobs
+        payload = []
+        timestamp = datetime.datetime.now(datetime.timezone.utc)
+
+        for event in events:
+            serialized = self.serializer.serialize_event(event)
+            payload.append({"blob": serialized})
+
+        # Store all events in a single create_event call
+        self.client.create_event(
+            memoryId=self.memory_id,
+            actorId=actor_id,
+            sessionId=session_id,
+            eventTimestamp=timestamp,
+            payload=payload,
+        )
+
+    def get_events(
+        self, session_id: str, actor_id: str, limit: int = 100
+    ) -> List[EventType]:
+        """Retrieve events from AgentCore Memory."""
+
+        if limit is not None and limit <= 0:
+            return []
+
+        all_events = []
+        next_token = None
+
+        while True:
+            params = {
+                "memoryId": self.memory_id,
+                "actorId": actor_id,
+                "sessionId": session_id,
+                "maxResults": 100,
+                "includePayloads": True,
+            }
+
+            if next_token:
+                params["nextToken"] = next_token
+
+            response = self.client.list_events(**params)
+
+            for event in response.get("events", []):
+                for payload_item in event.get("payload", []):
+                    blob = payload_item.get("blob")
+                    if blob:
+                        try:
+                            parsed_event = self.serializer.deserialize_event(blob)
+                            all_events.append(parsed_event)
+                        except EventDecodingError as e:
+                            logger.warning(f"Failed to decode event: {e}")
+
+            next_token = response.get("nextToken")
+            if not next_token or (limit is not None and len(all_events) >= limit):
+                break
+
+        return all_events
+
+    def delete_events(self, session_id: str, actor_id: str) -> None:
+        """Delete all events for a session."""
+        params = {
+            "memoryId": self.memory_id,
+            "actorId": actor_id,
+            "sessionId": session_id,
+            "maxResults": 100,
+            "includePayloads": False,
+        }
+
+        while True:
+            response = self.client.list_events(**params)
+            events = response.get("events", [])
+
+            if not events:
+                break
+
+            for event in events:
+                self.client.delete_event(
+                    memoryId=self.memory_id,
+                    sessionId=session_id,
+                    eventId=event["eventId"],
+                    actorId=actor_id,
+                )
+
+            next_token = response.get("nextToken")
+            if not next_token:
+                break
+            params["nextToken"] = next_token
+
+
+class EventProcessor:
+    """Processes events into checkpoint data structures."""
+
+    @staticmethod
+    def process_events(
+        events: List[EventType],
+    ) -> tuple[
+        Dict[str, CheckpointEvent],
+        Dict[str, List[WriteItem]],
+        Dict[tuple[str, str], Any],
+    ]:
+        """Process events into organized data structures."""
+        checkpoints = {}
+        writes_by_checkpoint = defaultdict(list)
+        channel_data_by_version = {}
+
+        for event in events:
+            if isinstance(event, CheckpointEvent):
+                checkpoints[event.checkpoint_id] = event
+
+            elif isinstance(event, WritesEvent):
+                writes_by_checkpoint[event.checkpoint_id].extend(event.writes)
+
+            elif isinstance(event, ChannelDataEvent):
+                if event.value != EMPTY_CHANNEL_VALUE:
+                    channel_data_by_version[(event.channel, event.version)] = (
+                        event.value
+                    )
+
+        return checkpoints, writes_by_checkpoint, channel_data_by_version
+
+    @staticmethod
+    def build_checkpoint_tuple(
+        checkpoint_event: CheckpointEvent,
+        writes: List[WriteItem],
+        channel_data: Dict[tuple[str, str], Any],
+        config: CheckpointerConfig,
+    ) -> CheckpointTuple:
+        """Build a CheckpointTuple from processed data."""
+        # Build pending writes
+        pending_writes = [
+            (write.task_id, write.channel, write.value) for write in writes
+        ]
+
+        # Build parent config
+        parent_config = None
+        if checkpoint_event.parent_checkpoint_id:
+            parent_config = {
+                "configurable": {
+                    "thread_id": config.thread_id,
+                    "actor_id": config.actor_id,
+                    "checkpoint_ns": config.checkpoint_ns,
+                    "checkpoint_id": checkpoint_event.parent_checkpoint_id,
+                }
+            }
+
+        # Build checkpoint with channel values
+        checkpoint = checkpoint_event.checkpoint_data.copy()
+        channel_values = {}
+
+        for channel, version in checkpoint.get("channel_versions", {}).items():
+            if (channel, version) in channel_data:
+                channel_values[channel] = channel_data[(channel, version)]
+
+        checkpoint["channel_values"] = channel_values
+
+        return CheckpointTuple(
+            config={
+                "configurable": {
+                    "thread_id": config.thread_id,
+                    "actor_id": config.actor_id,
+                    "checkpoint_ns": config.checkpoint_ns,
+                    "checkpoint_id": checkpoint_event.checkpoint_id,
+                }
+            },
+            checkpoint=checkpoint,
+            metadata=checkpoint_event.metadata,
+            parent_config=parent_config,
+            pending_writes=pending_writes,
+        )