Most "AI memory" demos are a vector store with a marketing label. You embed every message, cosine-search the top-k on the next turn, and call it memory. It works until turn 40, when the agent confidently tells a user their favorite color is blue because that's what came back highest-ranked — even though they corrected it three messages later.
Real memory isn't retrieval. It's deciding what's still true. That distinction is why four very different architectures — Mem0, Zep, LangChain's memory classes, and Letta (formerly MemGPT) — all claim the same territory but solve almost none of the same problems. Here's what each one actually does under the hood, where it breaks, and which one you should reach for.
LangChain memory: primitives, not a system
ConversationBufferMemory, ConversationSummaryMemory, ConversationKGMemory, VectorStoreRetrieverMemory — these are building blocks, not a memory service. You own the extraction logic, the storage schema, and every decision about what gets kept or discarded. ConversationSummaryMemory re-summarizes the whole history on every turn, which means cost and latency grow with conversation length even though the output size doesn't. ConversationKGMemory extracts triples but has no mechanism to invalidate a triple once a new fact contradicts it — old and new coexist in the graph, and retrieval has no way to prefer one.
