Is RAG Becoming Outdated Now That Long-Running Agents Like Claude Cowork Are Emerging?

This post analyzes the evolving landscape of AI agent memory, contrasting traditional RAG with emerging long-running agent capabilities like Claude Cowork. It details how Claude Cowork uses a hybrid approach of persistent user memory and on-demand conversational retrieval (semantic and keyword search) to maintain state across long sessions, differentiating it from RAG's knowledge-driven retrieval. The post also contrasts this with ChatGPT's more structured memory system (User Memory, Session Metadata, cross-session summaries). It highlights the challenges of implementing writable agent memory, including deciding what to remember, how to store and tier memory, and how to write fast without breaking retrieval. The core technical contribution is the analysis of these memory management strategies and their implications for vector databases in AI agent development.