Context and memory solve different problems. A context window controls how much information a model can inspect during an inference or conversation. Agent memory controls which facts and decisions should survive beyond that conversation and become useful to a future agent.

A larger context window can reduce repetition inside a session, but capacity does not perform judgment. It does not decide which decision remains authoritative, which preference was superseded, which failed approach should be avoided, or which fact has become stale. Continuity requires selection, maintenance, provenance, retrieval, and forgetting.

This distinction matters most when models, coding-agent harnesses, projects, or machines change. The intelligence consuming memory can be replaceable. Durable user state should remain portable and inspectable underneath it.