T10 Erasure resurrection
The user asked for a record to be erased and the tool erased it. The attacker restores the copy of the store he took before the erasure, sidecar files and all.
What stops it
A head that moved past the erasure and is kept where the restore cannot reach it: a witness on another machine, an anchor held off the store. The restored copy is then behind the head, and the read refuses.
What does not
A ledger of what must be present, including one that catches a whole-store rollback: after the restore, every record it ever recorded is present again. T10 is not yet placed on the levels ladder; the draft that defines the levels has not placed it.
Today, measured 2026-10-19
- rejected on read
- Recallspection ExactMemory, anchored, anchor off the store, on a separate volume or a remote service
- accepted
- LangGraph SqliteSaver, OpenAI Agents SDK SQLiteSession, LlamaIndex Memory, SQLAlchemy chat store, Mem0 local Qdrant store, langgraph-ledger over SqliteSaver, hash-chained ledger, verify_thread audit, Recallspection ExactMemory, anchor beside the store on the same disk, the default the product ships
agmi-check --adapter agmi.adapters.langgraph_sqlite:LangGraphSqliteAdapter # runs T1 to T8; this is T10