SQL-backed memory service for the Google Agent Development Kit (ADK).

Drop-in durable replacement for InMemoryMemoryService. Works with any async SQLAlchemy dialect: SQLite, PostgreSQL, MySQL/MariaDB, and more.

Listed in the official ADK integrations catalog.

Why

ADK ships an in-memory memory service for development and Vertex AI Memory Bank for production on GCP. If you want durable memory on your own Postgres / SQLite / MySQL, there was no built-in option. This package fills that gap with the same BaseMemoryService contract, so you can swap it in without changing agent code.

Install

pip install adk-database-memory[sqlite]       # SQLite (via aiosqlite)
pip install adk-database-memory[postgres]     # PostgreSQL (via asyncpg)
pip install adk-database-memory[mysql]        # MySQL / MariaDB (via aiomysql)

The base install does not pull any database driver. Pick the extra that matches your backend, or install your own async SQLAlchemy driver separately.

Examples

Three runnable scripts in examples/:

quickstart_sqlite.py - smallest end-to-end usage
with_runner.py - plug into an ADK Runner with the load_memory tool
multi_user_search.py - per-user isolation demo

Quick start

import asyncio
from adk_database_memory import DatabaseMemoryService
async def main():
    # SQLite, zero-config
    async with DatabaseMemoryService("sqlite+aiosqlite:///memory.db") as memory:
        await memory.add_session_to_memory(session)
        result = await memory.search_memory(
            app_name="my_app",
            user_id="u1",
            query="what did we decide about the pricing model?",
        )
        for entry in result.memories:
            print(entry.author, entry.timestamp, entry.content)
asyncio.run(main())

With an ADK `Runner`

from google.adk.runners import Runner
from adk_database_memory import DatabaseMemoryService
memory = DatabaseMemoryService(
    "postgresql+asyncpg://user:pass@localhost:5432/agentdb"
)
runner = Runner(
    app_name="my_app",
    agent=my_agent,
    memory_service=memory,
    session_service=session_service,
)

Supported backends

Backend	URL example	Extra
SQLite	`sqlite+aiosqlite:///memory.db`	`[sqlite]`
SQLite (in-memory)	`sqlite+aiosqlite:///:memory:`	`[sqlite]`
PostgreSQL	`postgresql+asyncpg://user:pass@host/db`	`[postgres]`
MySQL / MariaDB	`mysql+aiomysql://user:pass@host/db`	`[mysql]`
Any async SQLAlchemy dialect	depends on driver	bring your own

API

The service implements google.adk.memory.base_memory_service.BaseMemoryService, so it exposes the same three methods used everywhere else in ADK:

add_session_to_memory(session) - index every event of a session.
add_events_to_memory(app_name, user_id, events, ...) - index a delta slice of events (useful for streaming ingestion).
search_memory(app_name, user_id, query) - return MemoryEntrys whose indexed keywords overlap with the query, scoped to the given app and user.

Constructor:

DatabaseMemoryService(
    db_url: str,
    *,
    stop_words: set[str] | None = None,   # override the default English stop-words list
    **engine_kwargs,                       # forwarded to create_async_engine
)

Lifecycle:

async with DatabaseMemoryService(db_url) as memory:
    ...
# equivalent to:
memory = DatabaseMemoryService(db_url)
try:
    ...
finally:
    await memory.close()

How it works

On first write, a single table adk_memory_entries is created (lazy, with an async double-checked lock) with an index on (app_name, user_id).
Each event's text content is lower-cased, tokenized ([A-Za-z]+), and filtered against the stop-words set to produce a keyword bag.
Search extracts keywords from the query the same way and returns rows where the bags overlap, scoped by app_name and user_id, then de-duplicates on (author, text, timestamp).
JSON content is stored as JSONB on PostgreSQL, LONGTEXT on MySQL, and TEXT on SQLite/others via a DynamicJSON type decorator.

This is intentionally the same keyword-matching approach as the in-memory and Firestore memory services in ADK; it is a durable, zero-infra starting point, not a semantic retriever. If you need embedding-based recall, pair this package with Vertex AI Memory Bank or a vector store.

Development

git clone https://github.com/anmolg1997/adk-database-memory
cd adk-database-memory
pip install -e ".[dev]"
pytest
ruff check .
mypy src

License

Apache 2.0 - see LICENSE.

Official ADK docs listing - integration page on google.github.io/adk-docs.
google/adk-python - core Agent Development Kit.
ADK Memory overview - background on how ADK uses memory services.
FirestoreSessionService (PR #5088) - the sibling Firestore session service this memory service's keyword-index approach mirrors.

Why