Make ADK application agent use Mem0 to persist information
Overview
This tutorial details the process of building and deploying a Google Agent Development Kit (ADK) application that utilizes mem0 to retain information across different sessions. It is based on mem0 Google Agent Development Kit example .
It uses ADK agent tools to invoke mem0 client that has to save or search information from the vector database that is hosted on CloudSQL. More information on the used agent can be found in the source code.
The tutorial will cover:
- Setting up your Google Cloud environment.
- Building a container image for your agent.
- Deploying the agent to a GKE cluster.
- Testing your deployed agent.
Before you begin
Ensure you have the following tools installed on your workstation
If you previously installed the gcloud CLI, get the latest version by running:
gcloud components update
Ensure that you are signed in using the gcloud CLI tool. Run the following command:
gcloud auth application-default login
Infrastructure Setup
Clone the repository
Clone the repository with our guides and cd to the tutorial directory by running these commands:
git clone https://github.com/ai-on-gke/tutorials-and-examples.git
cd tutorials-and-examples/adk/memory/
Filesystem structure
├── adk-app # ADK application source
│ ├── agent_with_memory
│ │ ├── agent.py
│ │ └── __init__.py
│ ├── Dockerfile
│ ├── main.py
│ └── requirements.txt
├── gen # This folder will contain all files that are generated by the terraform
└── terraform # The terraform config source
├── templates
│ └── agent-with-memory.yaml.tftpl # Template of the ADK app Kubernetes manifest
├── cloudsql.tf
├── example.tfvars
├── main.tf
├── network.tf
├── outputs.tf
├── providers.tf
├── variables.tf
└── versions.tf
Create cluster and other resources
In this section we will use Terraform to automate the creation of infrastructure resources. For more details how it is done please refer to the terraform config in the terraform/ folder. By default, the configuration provisions an Autopilot GKE cluster, but it can be changed to standard by setting autopilot_cluster = false.
It creates the following resources. For more information such as resource names and other details, please refer to the Terraform config:
- Service Accounts:
- Cluster IAM Service Account (derives name from a cluster name, e.g.
tf-gke-<cluster name>) – manages permissions for the GKE cluster.
- Cluster IAM Service Account (derives name from a cluster name, e.g.
- Artifact registry – stores container images for the application.
-
Go the the terraform directory:
cd terraform -
Specify the following values inside the
example.tfvarsfile (or make a separate copy):<PROJECT_ID>– replace with your project id (you can find it in the project settings).
Other values can be changed, if needed, but can be left with default values.
-
Init terraform modules:
terraform init -
Optionally run the plan command to view an execution plan:
terraform plan -var-file=example.tfvars -
Execute the plan:
terraform apply -var-file=example.tfvars -
Configure your kubectl context:
gcloud container clusters get-credentials $(terraform output -raw gke_cluster_name) --region $(terraform output -raw gke_cluster_location)
Deploy and Configure the Agent Application
-
Build image with our ADK application:
gcloud builds submit \ --tag $(terraform output -raw image_repository_full_name)/agent-with-memory:latest \ --project=$(terraform output -raw project_id) \ ../adk-app -
Create a secret with a Google API key:
kubectl create secret generic google-api-secret --from-literal=key="<YOUR_KEY>" -
Deploy the manifest for the ADK application:
kubectl apply -f ../gen/agent-with-memory.yaml -
Wait until the deployment is ready:
kubectl rollout status deployment/adk-agent
Test the ADK app and memory
-
Expose the webui port of the ADK app:
starYou can also use our Identity Aware Proxy guide to securely expose the app.
kubectl port-forward svc/adk-agent 8080:80 -
Make agent save something in memory, for example by prompting “I am planning to buy a car next week”:
You should see a successful call of the
save_memorytool, which has persisted information of your prompt in mem0. -
Verify saved information:
3.1. Create a new session:
3.2. Make agent search its memory, for example by prompting “What are my plans for next month?”:
starFirst use of the mem0 may take a few seconds for the database to be ready.
The agent should find previously saved information by using the
search_memorytool.
Cleaning up
-
Destroy the provisioned infrastructure.
terraform destroy -var-file=example.tfvars