Context is not memory
Treating the context window like a conversation history is the first mistake most teams make when building agents.
There's a conceptual error that shows up consistently in early-stage agent implementations: treating the context window as a memory system. It isn't one, and building as if it is produces systems that degrade in predictable ways.
Memory implies persistence, retrieval, and selective access. A context window has none of these properties. It's a fixed-size input buffer. Everything you put in it costs inference tokens. Nothing persists across calls unless you explicitly carry it over. The model doesn't 'remember' prior turns — it sees a document that contains prior turns, which is a different thing with different implications.
The practical consequence: context management is not an afterthought. Deciding what goes into the context window on each call — what's essential, what can be summarised, what can be retrieved on demand — is one of the most important architectural decisions in an agent system. It determines cost, latency, quality, and how gracefully the system handles long-running tasks.
Build context management as a first-class component. It belongs in your architecture diagram, not your prompt template.