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 TTL (default 15 days).
• Shared storage. Memory uses the kagent database (PostgreSQL), not a separate database.

Configuration#

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

Custom TTL#

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

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

How Memory Works#

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.

Viewing Memories#

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.

Memory Management via API#

The following API endpoints let you manage memories programmatically.

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.

Technical Details#

• Embedding vectors are normalized to 768 dimensions.
• Background TTL pruning runs periodically with a default retention of 15 days. Memories older than the retention period are automatically deleted.
• Memories include timestamps and source session references.