Prove Recall and Erasure

One database or SDK rarely solves every memory job cleanly. Maya's agent needs a semantic preference across tickets and checkpointed progress inside one ticket, so use Mem0 for approved durable facts and LangGraph persistence for transient workflow recovery.

def prepare_support_context(tenant_id: str, user_id: str, ticket_id: str, question: str):
    scope = f"{tenant_id}:{user_id}"
    facts = memory.search(question, user_id=scope)
    checkpoint_config = {"configurable": {"thread_id": f"{scope}:{ticket_id}"}}
    return {"facts": facts, "checkpoint_config": checkpoint_config}

This function makes two lifetimes explicit. Never promote checkpoint transcripts into profile memory without a separate approval path.

Retrieval quality is a product property, not a demo. Keep a small fixed evaluation set with expected relevant facts, expected absent facts, and a threshold you can regress-test.

def test_maya_memory_is_relevant_and_isolated():
    maya = memory.search("best channel for Maya update", user_id="acme:maya")
    assert any("email" in hit["memory"].lower() for hit in maya)

    stranger = memory.search("best channel for Maya update", user_id="other-tenant:maya")
    assert not any("email" in hit["memory"].lower() for hit in stranger)

Record backend version, fixture version, score, and failures. This proves this test case only; it does not prove universal safety or relevance.

Deletion is incomplete if one copy remains in a vector index, checkpoint store, cache, or log. Build an inventory of every store your design owns, then delete by tenant-qualified reference and verify no later search returns the fact.

from typing import Protocol

class MemoryBackend(Protocol):
    def search(self, query: str, *, user_id: str) -> list[dict]: ...
    def delete_user(self, *, user_id: str) -> None: ...

def erase_user(tenant_id: str, user_id: str):
    scope = f"{tenant_id}:{user_id}"
    memory_backend.delete_user(user_id=scope)
    checkpoint_backend.delete_thread_prefix(f"{scope}:")
    audit_log({"event": "memory.erase", "scope": scope})

erase_user("acme", "maya")
assert memory_backend.search("Maya email preference", user_id="acme:maya") == []

MemoryBackend is an adapter contract, not a claim that every provider exposes the same delete call. Map it to each selected provider's documented deletion API; backups, legal holds, and immutable audit logs can have separate retention rules. Explain those exceptions to users instead of promising immediate universal erasure.

Memory is user data with an AI-shaped retrieval path. Assign an owner, review write policy, measure retrieval, restrict tenant access, set retention, and keep deletion receipts.