Does AI actually need long-term memory, or is context window scaling enough?

I’ve been following the recent developments in LLMs, specifically the debate between Long-Term Memory (RAG/vector databases) and Massive Context Windows (like 1M+ tokens). It got me thinking: do models actually need a human-like memory system, or are we just overcomplicating things? The Argument for Long-Term Memory (RAG / Knowledge Graphs): Cost & Speed: Feeding millions of tokens into every prompt is computationally expensive and slow. Human-like Dynamics: Real intelligence doesn't reload its entire life history every time it speaks. It retrieves relevant memories, updates beliefs, and forgets irrelevancies. Personalization: A structured long-term memory allows an AI agent to truly grow with the user over months or years without running out of space. The Argument for Unlimited Context Windows: No Information Loss: RAG often misses nuances or retrieves the wrong chunks. A massive context window reads everything directly. Simplicity: Why build complex retrieval pipelines, embeddings, and vector DBs if a model can just process the entire document/history natively? Reasoning Capabilities: High-attention context allows the model to draw connections across distant pieces of information that retrieval algorithms might separate. My Question to the Community: Are we moving toward a future where "memory" is just an optimized, persistent context window, or will dedicated memory architectures (episodic, semantic, procedural) always be necessary for true AI agents? What’s your setup for handling long-term interactions right now?