Match Memory Mechanism to Job

Mem0 gives semantic fact recall, but Maya's agent may also need to resume a half-finished refund workflow. A checkpoint records execution state; it should not become a permanent profile.

LangGraph persistence stores checkpoints by thread. Its pros: explicit resumable graph state and time travel. Cons: application-managed schema, storage, and lifecycle; it does not automatically decide which cross-thread fact matters.

from langgraph.checkpoint.memory import InMemorySaver

checkpointer = InMemorySaver()
config = {"configurable": {"thread_id": "acme:maya:ticket-42"}}
# Compile graph with checkpointer=checkpointer, then invoke with config.

Use a production checkpointer backend; in-memory state disappears when this process exits.

Different memory APIs own different units of state. This table is a selection aid, not a benchmark.

A vendor's storage abstraction cannot decide who may see a fact. Build a small application boundary that verifies tenant identity, calls chosen backend, and logs an opaque memory reference rather than copying sensitive text into logs.

type MemoryRequest = { tenantId: string; userId: string; query: string };

export async function recall(req: MemoryRequest) {
  const namespace = `${req.tenantId}:${req.userId}`;
  const hits = await memoryBackend.search({ query: req.query, namespace });
  auditLog({ event: "memory.search", namespace, hitCount: hits.length });
  return hits;
}

memoryBackend is an adapter you own. Validate caller authorization before this function; string namespaces are a routing convention, not an access-control system.

Choice becomes clearer when based on evidence: Maya's team needs profile recall and resumable tickets now, not a full temporal graph. Choose Mem0 for curated facts plus LangGraph checkpoints for active ticket recovery; reassess Zep only if time-based relationship questions become product requirements.