Most production agent stacks already have the raw ingredients for memory: a LangGraph runtime, a MongoDB checkpoint store, and session state at the edge. So why reach for a hosted memory service? This talk walks through that exact decision using a real back end: LangGraph-driven orchestration, MongoDB checkpointers, and per-session conversation storage, and asks whether the in-house pieces are enough, or whether off-the-shelf memory platforms such as Zep, Mem0, LangMem, and Letta earn their keep.The session compares the two paths across the dimensions that matter once an agent leaves the demo: data residency and personally identifiable information (PII) handling, retrieval latency under real traffic, evaluations and drift detection, schema evolution, and the long-tail maintenance cost of summarization, decay, and conflict resolution. Building in-house wins on control, tenancy, and integration with existing authentication and audit trails, but absorbs every line of recall, ranking, and forgetting logic. Off-the-shelf solutions win on time-to-value and battle-tested retrieval, but introduce a third party in the data path, a new dependency to monitor, and opinionated abstractions that may not survive contact with a multi-tenant software-as-a-service (SaaS) product.Attendees will leave with a concrete decision framework: when a LangGraph checkpoint plus a MongoDB collection is genuinely sufficient, when a vector-backed semantic layer earns its place, and the specific signals — PII surface area, cross-session recall needs, and evaluation maturity — that should push a team one way or the other.