bedrock-memorysync

MemorySync reference architecture for AWS Bedrock agents — two runnable patterns:

1. MemoryConverseLoop — a memory-augmented bedrock-runtime.converse loop: recall injected as a system block under a hard budget, and the user's message sent to fact extraction after each reply (only durable facts become memories; assistant replies are not stored).
2. Bedrock Agents action group — a Lambda handler + function schema that gives any Bedrock Agent four memory functions. No memory vendor (including Mem0) ships a Bedrock Agents action-group integration — this is the first.

pip install bedrock-memorysync

Pattern 1 — the Converse loop

import boto3
from bedrock_memorysync import MemoryConverseLoop

loop = MemoryConverseLoop(
    bedrock_client=boto3.client("bedrock-runtime"),
    model_id="anthropic.claude-3-haiku-20240307-v1:0",
    user_id="customer-42",                       # required
    system_prompt="You are a helpful travel assistant.",
)
loop.chat("I always prefer window seats on long flights")
# … a new session, days later:
loop.new_session("support-2")
loop.chat("which seat should I book for the Oslo flight?")   # remembers

• Recall runs under recall_timeout (default 1.2 s) and fails open to a memory-free turn — memory can never delay or break the model call.
• After the reply, the user's message goes to fact extraction with a deterministic seed: MemorySync stores the durable facts in it (preferences, personal details, decisions), not the text itself, and recognises a retried turn so it is processed once. Assistant replies are not stored as memories, so the loop does not send them.
• You own the boto3 client (region, credentials, retry policy) — tests drive it with botocore's Stubber.

Pattern 2 — Bedrock Agents action group

Deploy the Lambda (SAM):

sam deploy --guided --parameter-overrides MemorySyncApiKey=ms_...

Attach to your agent (function-details schema, importable):

from bedrock_memorysync import FUNCTION_SCHEMA

bedrock_agent.create_agent_action_group(
    agentId=agent_id, agentVersion="DRAFT",
    actionGroupName="memorysync-memory",
    actionGroupExecutor={"lambda": function_arn},
    functionSchema={"functions": FUNCTION_SCHEMA},
)

remember tells the agent what happened: status: "accepted" with stored_as: "facts" when the text went to fact extraction, or status: "skipped" with the reason in processing_status and message — an assistant turn (not stored as a memory), nothing worth remembering, the monthly quota, or already_exists: true when the same text was already sent. It never reports a memory id: facts are extracted asynchronously and get their own ids.

There is deliberately no delete-everything function — a model-reachable account wipe is a data-loss foot-gun.

Identity ladder (multi-tenant safe): sessionAttributes.memorysync_user_id (set by your app when invoking the agent) → a user_id parameter → the Bedrock sessionId. Every failure returns a structured responseState: FAILURE body the agent can react to — never an unhandled Lambda error surfaced as a wall of text.

Why not the alternatives?

Tests

pip install -e . boto3 pytest
pytest tests -q   # 41 checks

The suite drives the REAL boto3 client under botocore's Stubber (system-block injection asserted verbatim) and the Lambda handler with the documented Bedrock Agents event shapes — including the identity ladder, idempotent retries, FAILURE response states, and quota modes.

License

MIT © MemorySync.