Google ADK rewind_async still leaks undone turns into long-term memory

Disclosure: I work on MemorySync (I'm Mohammed, the founder). I drafted this page with AI assistance and checked it against the source issue, the fix commit and the MemorySync Google ADK guide on Oct 28, 2026. I didn't run any code for this page.

The question

google/adk-python#7570 (opened Oct 16 2026) reports that after Runner.rewind_async, calling add_session_to_memory still stores the turns the rewind took back. Rewind keeps every event in the session and appends a rewind event; prompt building filters through _apply_rewinds, but the memory services named in the issue (InMemoryMemoryService, SqliteMemoryService, VertexAiMemoryBankService, VertexAiRagMemoryService, FirestoreMemoryService) read session.events directly. So if a user takes back a wrong address or something private, load_memory or preload_memory can bring it back in a later session.

Status (checked 2026-10-28): closed as completed on Oct 28, 2026 (05:13 IST). PR #7571 ("keep turns a rewind took back out of long-term memory") landed on main as commit 2591397 (the PR itself shows as closed rather than merged, because the change was imported through Google's Copybara sync). The commit changes the in-memory, SQLite, Firestore, Vertex AI Memory Bank and Vertex AI RAG memory services. The latest ADK release, v2.13.0 (Oct 23), predates it, so the fix isn't in a tagged release yet.

What to do until the fix is released

Not run for this page.

• Don't ingest a rewound session as-is. If your app uses rewind, either skip add_session_to_memory for sessions that contain a rewind event, or build a filtered copy of the session that holds only live events before ingesting.
• Ingest at the end, not per turn. On v2.13.0 and earlier, memory written before a rewind isn't removed by the rewind; you'd have to delete it from the memory backend yourself. Per the fix commit, after upgrading, add_session_to_memory on the SQLite and Firestore services also deletes stored entries for turns a rewind took back.
• Watch for the first ADK release after v2.13.0 that includes commit 2591397, and re-check your memory service after upgrading ADK.

Common questions

Is it fixed?
On main, yes: issue #7570 was closed on Oct 28, 2026 by commit 2591397 (PR #7571). As of v2.13.0, the latest release when I checked, the fix isn't in a tagged release.

Does the fix delete memories that are already stored?
Per the commit message, on add_session_to_memory the SQLite and Firestore services now also delete stored entries for events a rewind took back. Entries for events that aren't in session.events at all, such as turns outside a session loaded with num_recent_events, are kept.

Where MemorySync fits

The MemorySync Google ADK guide documents a MemorySyncMemoryService whose add_session_to_memory sends the user's messages from session.events. The guide doesn't cover rewinds, so filter rewound events out of the session yourself before calling it. What MemorySync adds is a delete-by-ID operation scoped to a user, which helps if an undone fact was already stored. To see save and recall in your browser first, try the MemorySync playground.

Sources

• google/adk-python#7570, "After Runner.rewind_async, add_session_to_memory still stores the rewound turns in long-term memory" (read 2026-10-28)
• google/adk-python PR #7571 (closed; its change landed on main as 2591397; issue #7570 closed 2026-10-28; read 2026-10-28)
• ADK docs: rewind, https://adk.dev/sessions/session/rewind/
• MemorySync Google ADK guide: docs.memorysync.io/guides/google-adk