june-adk 0.1.0

Give your Google ADK agent a June memory — a local-first MemoryService backed by a searchable, cited June knowledge canvas, plus a backfill CLI for existing session databases.

june-adk

Give your Google ADK agent a June memory. ADK already defines a pluggable memory seam — BaseMemoryService — but its production implementation sends your agent's conversations to Vertex AI RAG in Google's cloud. june-adk implements the same seam against Junê instead: your agent's history becomes a local-first, searchable, cited knowledge graph, canvas-fenced like everything else in June.

your ADK agent ──▶ JuneMemoryService ──▶ June canvas (graph · search · citations)
                   (ADK's own seam)       local by default — your machine, your data

The live seam — one object in your Runner

pip install june-adk

from google.adk.runners import Runner
from june_adk import JuneMemoryService

memory = JuneMemoryService(
    june_url="http://127.0.0.1:8799",    # the Junê desktop app's local engine
    api_key="local",
    canvas="agent-memory",
    state_file=".june_adk_pushed.json",  # dedupe ledger, survives restarts
)
runner = Runner(agent=agent, app_name="my-app",
                session_service=session_service, memory_service=memory)

That's the whole integration. ADK's own memory tools (load_memory, preload-memory) now read June, and add_session_to_memory / add_events_to_memory write it. JuneMemoryService.from_env() builds from the same JUNE_BASE_URL / JUNE_API_KEY / JUNE_CANVAS env vars that june-mcp and june-langgraph use.

Contract fidelity, verified against the installed google-adk (not docs):

• add_session_to_memory — ADK says a session "may be added multiple times during its lifetime"; we honor that with per-event delta dedup, so repeated adds never duplicate and only new events ship.
• add_events_to_memory — the incremental path, treated as a delta exactly as the base class demands.
• search_memory → SearchMemoryResponse of MemoryEntry items, with June provenance (score, node id) in custom_metadata.
• Fail-soft by default: a memory write must never crash the agent — June being down logs one line and retries on the next add. Pass strict=True to raise instead. All HTTP runs in a worker thread; the event loop is never blocked.

The backfill — history your app already saved

If your app persists sessions with ADK's DatabaseSessionService, the history that already exists can be brought into June without touching your code:

pip install "june-adk[db]"
june-adk sync --db sessions.db          # one incremental catch-up pass
june-adk watch --db sessions.db         # keep syncing (Ctrl+C to stop)
june-adk doctor --db sessions.db        # PASS/FAIL every part of the setup

Discovery is metadata-only SQL (which app/user scopes exist); content is always read through ADK's own DatabaseSessionService, so ADK deserializes its events — never us. Bare file paths and sqlite:// URLs are quietly upgraded to the async driver ADK 2.x requires. Server databases (Postgres, MySQL) work too — pass --app and --user. Every pass is incremental via the same ledger the live seam uses; re-running never duplicates.

What you get on the June side

The canvas behaves like any other: search it, ask cited questions over it in the Junê app, and point june-mcp at the same canvas so Claude (or any MCP agent) can answer over your ADK agent's entire history. One memory, every agent.

Part of a family: june-langgraph does the same for LangGraph checkpointers.

License

MIT. The Junê engine itself is a separate, closed-source product — this connector is the open part, by design.