A thought-provoking Chinese engineering blog post challenges the prevailing approach to building long-term memory for AI agents. The author argues that most current implementations treat agent memory as a simple RAG pipeline: chunking historical messages, generating embeddings, storing them in a vector database, and retrieving top-K results on demand. While this works for MVPs, it fundamentally fails to address the needs of production-grade agents that require evidence tracking, version control, and lifecycle management of memories. The article proposes a more robust architecture where agent memory is treated as a first-class data system with its own schema, versioning, and consistency guarantees. This includes maintaining provenance of each memory entry, supporting updates and deletions with proper history, and implementing garbage collection for stale or irrelevant memories. The author also discusses trade-offs between different storage backends (relational, graph, and vector) and suggests a hybrid approach. This is a significant signal for the AI engineering community, as it identifies a critical gap in current agent architectures and offers a concrete direction for improvement.
This article argues that long-term agent memory should not be treated as a lightweight RAG system. Instead, it proposes a data system that maintains evidence, versioning, and lifecycle management for agent memories, going beyond simple vector similarity search.