adk-itsuki

Itsuki memory for Google ADK: a native BaseMemoryService, plus the automatic capture ADK does not ship.

Status: unpublished. This package has not been released to PyPI. The install command below will not work yet.

Why the plugin exists

ADK has a memory interface but nothing that fills it. Searching the framework for add_session_to_memory finds exactly two call sites: a convenience method your own callback can call, and an HTTP endpoint a client must invoke. The dev UI never calls either. So an ADK app gets retrieval for free and remembers nothing unless the application writes the plumbing.

ItsukiMemoryPlugin is that plumbing, placed where it fires once per invocation rather than once per agent.

Requirements

The floor is 2.5.0 for a specific reason: it is the first release whose normal run path invokes after_run_callback under a run_error guard and exposes on_run_error_callback. Below that, automatic capture and identity cleanup cannot work. The ceiling is the highest minor actually tested, because ADK ships breaking changes in minor releases.

Use

from google.adk.agents import LlmAgent
from google.adk.apps.app import App
from google.adk.runners import Runner
from google.adk.tools import preload_memory
from adk_itsuki import ItsukiMemoryService, ItsukiMemoryPlugin

memory = ItsukiMemoryService()          # reads ITSUKI_API_KEY
agent = LlmAgent(name="assistant", model="gemini-2.0-flash", tools=[preload_memory])

app = App(name="my_app", root_agent=agent, plugins=[ItsukiMemoryPlugin(service=memory)])
runner = Runner(app=app, app_name="my_app", session_service=..., memory_service=memory)

preload_memory gives you recall before every model call; the plugin gives you capture after every settled turn.

The itsuki:// URI route is partial

# services.py, next to your agent
import adk_itsuki
adk_itsuki.register()

adk web --memory_service_uri=itsuki://

This lets the CLI and dev UI construct the service. It does not attach preload_memory or the capture plugin — ADK has no mechanism for a service to do that. Treat it as a partial installation, not an automatic one.

What gets remembered

Only the user's message and the root agent's final answer. Specifically excluded: tool calls and their results, code execution, inline and file data, sub-agent intermediate output, partial streaming events, and anything from an invocation that errored or is still waiting on a long-running tool.

Attribution is by author name rather than by branch, because a SequentialAgent child can keep the root branch. The root agent's name is recorded once per invocation in session state (content-free) so that a restart, a delta import and an explicit re-import all reach the same verdict. An invocation with no marker is skipped rather than guessed at.

Honest limits

• Capture is durable to process life, not across restarts. Pending chunks live in memory. Recovery is an explicit add_session_to_memory(session), which regenerates identical idempotency keys, so re-importing costs nothing and duplicates nothing.
• A root SequentialAgent's children are not captured unless they author under the root's name. That is deliberate: guessing which sub-agent output a person actually saw would be worse than remembering less.
• add_memory is not implemented. Direct memory writes are a surface we have not audited yet.
• user_id is whatever your application passes. Hashing it prevents encoding and delimiter attacks; it cannot prevent impersonation. Bind user_id to an authenticated principal in your own auth layer.
• ADK 1.x is not supported.

Licence

Apache-2.0.