Use OpenSearch storage

The opensearch backend uses OpenSearch as the vector store. Memory text is embedded by an embedding model, written to OpenSearch with a per-user index (the actual index name is {index}_{user_id}), and retrieved by similarity. It is the default backend for LongTermMemory.

When to use

You already run an OpenSearch cluster and want self-hosted vector search;
You need persistence, sharing across processes and instances, and full control over indexing and retrieval.

Dependencies

pip install "veadk-python[extensions]"

Usage

Before running, provide a reachable OpenSearch cluster, an account allowed to create and access indexes, a valid CA certificate, and the connection and embedding variables below. Dimensions must match an existing index; use a new index and reinsert memories when changing dimensions

from veadk import Agent
from veadk.memory.long_term_memory import LongTermMemory

# index must be lowercase, only a-z0-9_-., and not start with _ or -
ltm = LongTermMemory(backend="opensearch", index="ltm_demo")

agent = Agent(
    name="demo",
    instruction="Answer the user; when needed, use the `load_memory` tool to recall past conversations.",
    long_term_memory=ltm,
)

You can also pass connection and embedding config explicitly via backend_config:

import os

from veadk.memory.long_term_memory import LongTermMemory
from veadk.configs.database_configs import OpensearchConfig

ltm = LongTermMemory(
    backend="opensearch",
    index="ltm_demo",
    backend_config={
        "index": "ltm_demo",
        "opensearch_config": OpensearchConfig(
            host="localhost",
            port=9200,
            username="admin",
            password=os.environ["DATABASE_OPENSEARCH_PASSWORD"],
            use_ssl=True,
            cert_path="/path/to/ca.pem",
        ),
    },
)

Parameters

LongTermMemory parameters

Parameter	Type	Default	Description
`backend`	`str \| BaseLongTermMemoryBackend`	`"opensearch"`	Set to `opensearch` for this page, or pass a configured backend instance
`backend_config`	`dict`	`{}`	Backend settings; a supplied backend instance takes precedence
`index`	`str`	`""`	Without backend_config, resolves from index, app_name, then default_app; supply a nonempty index with a configuration dictionary
`app_name`	`str`	`""`	Fallback for index; the actual user comes from the saved Session or search arguments
`top_k`	`int`	`5`	Number of retrieved chunks; use a positive integer
`user_id`	`str`	`""`	Deprecated; does not select the runtime user

Constructor parameters

backend_config supports the following settings:

Parameter	Type	Default	Description
`index`	`str`	No default; provided by `LongTermMemory`	Memory index name; must follow OpenSearch naming rules.
`opensearch_config`	`OpensearchConfig`	Read automatically from `DATABASE_OPENSEARCH_*` env vars	OpenSearch connection config.
`embedding_config`	`EmbeddingModelConfig`	Read automatically from `MODEL_EMBEDDING_*` env vars	Embedding model config.

OpenSearch connection config

opensearch_config is an OpensearchConfig with env prefix DATABASE_OPENSEARCH_:

Field	Env var	Type	Default	Description
`host`	`DATABASE_OPENSEARCH_HOST`	`str`	`""`	OpenSearch host.
`port`	`DATABASE_OPENSEARCH_PORT`	`int`	`9200`	Port.
`use_ssl`	`DATABASE_OPENSEARCH_USE_SSL`	`bool`	`true`	Whether to enable SSL.
`username`	`DATABASE_OPENSEARCH_USERNAME`	`str`	`""`	Username.
`password`	`DATABASE_OPENSEARCH_PASSWORD`	`str`	`""`	Password.
`cert_path`	`DATABASE_OPENSEARCH_CERT_PATH`	`str`	`""`	CA certificate path. When empty, certificates are not verified — a security risk that triggers a warning.
`secret_token`	`DATABASE_OPENSEARCH_SECRET_TOKEN`	`str`	`""`	Not used by current OpenSearch long-term memory connections

Embedding config

embedding_config is an EmbeddingModelConfig with env prefix MODEL_EMBEDDING_:

Field	Env var	Type	Default	Description
`name`	`MODEL_EMBEDDING_NAME`	`str`	`doubao-embedding-vision-250615`	Embedding model name.
`dim`	`MODEL_EMBEDDING_DIM`	`int`	`2048`	Embedding vector dimension; used to build the index’s vector field.
`api_base`	`MODEL_EMBEDDING_API_BASE`	`str`	`https://ark.cn-beijing.volces.com/api/v3/`	API base of the embedding service.
`api_key`	`MODEL_EMBEDDING_API_KEY`	`str`	Falls back to `MODEL_AGENT_API_KEY`, then an auto-fetched Ark token	Key for accessing the embedding service.

Environment variables

# OpenSearch connection
export DATABASE_OPENSEARCH_HOST="localhost"
export DATABASE_OPENSEARCH_PORT=9200
export DATABASE_OPENSEARCH_USERNAME="admin"
export DATABASE_OPENSEARCH_PASSWORD="admin"
export DATABASE_OPENSEARCH_USE_SSL=true
export DATABASE_OPENSEARCH_CERT_PATH="/path/to/ca.pem"

# Embedding model
export MODEL_EMBEDDING_NAME="doubao-embedding-vision-250615"
export MODEL_EMBEDDING_DIM=2048
export MODEL_EMBEDDING_API_KEY="your-ark-api-key"

index must follow OpenSearch naming rules: all lowercase, only a-z0-9_-., and not starting with _ or -; otherwise initialization fails. In production, set cert_path to enable certificate verification and avoid security risks.

Verify writes and retrieval

After configuring the dependencies and credentials on this page, run this standalone example. It saves user text and searches for that user directly without calling a conversation model

import asyncio

from google.adk.events import Event
from google.adk.sessions import Session
from google.genai import types
from veadk.memory.long_term_memory import LongTermMemory

async def main():
    memory = LongTermMemory(backend="opensearch", index="ltm_demo")
    session = Session(
        id="memory_check", app_name="ltm_demo", user_id="user_42",
        events=[Event(author="user", content=types.Content(
            role="user", parts=[types.Part(text="My preferred language is Chinese")]
        ))],
    )
    await memory.add_session_to_memory(session)
    result = await memory.search_memory(
        app_name="ltm_demo", user_id="user_42", query="preferred language"
    )
    for entry in result.memories:
        print(entry.content)

asyncio.run(main())

Results should contain the saved language preference. Managed services may extract memories asynchronously, so a completed write does not guarantee immediate retrieval. An empty result can also indicate permission, network, or service failure; check error logs and service records. The save method does not return a success Boolean The final index includes user_id, so user IDs must also meet the index character rules. The same index and user_id access the same memory; search_memory(app_name=...) does not add another index boundary

Overview

Get started

Core components

Observability

Extensions

Frontend

Security

Deployment

When to use

Dependencies

Usage

Parameters

LongTermMemory parameters

Constructor parameters

OpenSearch connection config

Embedding config

Environment variables

Verify writes and retrieval

​When to use

​Dependencies

​Usage

​Parameters

​LongTermMemory parameters

​Constructor parameters

​OpenSearch connection config

​Embedding config

​Environment variables

​Verify writes and retrieval

When to use

Dependencies

Usage

Parameters

LongTermMemory parameters

Constructor parameters

OpenSearch connection config

Embedding config

Environment variables

Verify writes and retrieval