Agent Memory#

With agent memory, your agents can automatically save and retrieve relevant context across conversations using vector similarity search. Memory is built on top of the Google ADK memory implementation.

Overview#

Agent memory provides the following capabilities.

• Vector-backed. A basic vector store uses embedding models to encode memories as 768-dimensional vectors.
• Searchable. Agents retrieve relevant memories via cosine similarity.
• Automatic. Agents extract and save user intent, key learnings, and preferences without explicit user action.
• Time-bounded. Memories expire after a configurable time to live (TTL), which defaults to 15 days.
• Metadata. Memories include timestamps and source session references.
• Shared storage. Memory uses the kagent database (PostgreSQL), not a separate database.

Configure memory#

Install kagent with Postgres#

To use memory, you must install kagent with a Postgres database that has the pgvector extension installed and vector enabled. For more information, see the Database configuration section.

Example Helm configuration for your own external Postgres:

database:
  postgres:
    urlFile: /var/secrets/db-url
    vectorEnabled: true
    bundled:
      enabled: false
controller:
  volumes:
    - name: db-secret
      secret:
        secretName: my-postgres-url-secret
  volumeMounts:
    - name: db-secret
      mountPath: /var/secrets
      readOnly: true

Enable memory on an agent#

To enable memory, set the memory field on the declarative agent spec, as shown in the following YAML example. The modelConfig field references a ModelConfig object whose embedding provider generates memory vectors.

You can also configure memory in the kagent UI when you create or edit an agent. In the UI, select an embedding model and set the memory TTL.

apiVersion: kagent.dev/v1alpha2
kind: Agent
metadata:
  name: memory-agent
  namespace: kagent
spec:
  type: Declarative
  declarative:
    modelConfig: default-model-config
    systemMessage: "You are a helpful assistant with long-term memory."
    memory:
      modelConfig: default-model-config  # References the embedding provider

The embedding ModelConfig does not have to use the same provider as the agent's main LLM. Kagent supports several embedding providers, including Amazon Bedrock as shown in the following section.

ModelConfig example: Amazon Bedrock#

To use Amazon Bedrock Titan embedding models, create a ModelConfig with provider: Bedrock.

The Bedrock provider uses the standard AWS credential chain, so no API key secret is required. The agent's pod must have AWS credentials with the bedrock:InvokeModel permission for the chosen model. On Kubernetes, the recommended setup is EKS IRSA on the agent ServiceAccount.

apiVersion: kagent.dev/v1alpha2
kind: ModelConfig
metadata:
  name: bedrock-embed
  namespace: kagent
spec:
  provider: Bedrock
  model: amazon.titan-embed-text-v2:0
  bedrock:
    region: us-east-1

Reference the embedding ModelConfig from the agent's memory.modelConfig field:

spec:
  declarative:
    memory:
      modelConfig: bedrock-embed

Optional: Custom TTL#

To change the default memory retention period of 15 days, set the ttlDays field.

memory:
  modelConfig: default-model-config
  ttlDays: 30  # Memories expire after 30 days instead of the default 15

Work with memories#

Review the following sections to understand what your agents can do with memory.

Automatic save cycle#

1. The agent processes user messages normally.
2. Every 5th user message, the agent automatically extracts key information such as user intent, key learnings, and preferences.
3. The agent summarizes extracted memories and encodes them as embedding vectors.
4. The agent stores the vectors in the database with metadata and timestamps.

Memory retrieval (prefetch)#

Before generating a response, the agent performs the following steps.

1. Encodes the current user message as an embedding vector.
2. Searches stored memories by cosine similarity.
3. Injects the most relevant memories into the agent's context.

Memory tools#

When you enable memory, the agent receives three additional tools.

You can also instruct the agent to use save_memory or load_memory explicitly during a conversation.

View memories in the UI#

In the kagent UI, you can view the memories that an agent has saved. With saved memories, you can inspect what the agent has learned and retained from past interactions.

Manage memories via API#

The following API endpoints let you manage memories programmatically.

Add a memory:

Add memories in batch:

Search memories by cosine similarity:

List memories:

Delete all memories for an agent:

Known limitations#

• No per-memory deletion. You can delete all memories for an agent, but you cannot delete individual memory entries.
• No cross-agent memory sharing. Each agent has its own isolated memory store. You cannot share memories across agents.
• Not pluggable. Memory is built on the Google ADK memory implementation and cannot be swapped for an alternative memory solution (such as Cognee). However, if an alternative memory solution exposes an MCP server, you can add it as a tool and instruct the agent to use it instead of the built-in memory.