Will million-token context windows make memory obsolete?

No. A million-token window means you CAN load more context. It doesn't mean you HAVE good context to load. Structured memory provides organized, prioritized, distilled knowledge. Raw conversation dumps provide noise with signal mixed in. The window size doesn't change this fundamental distinction.

Can I use the context window for short-term memory within a session?

Absolutely — that's exactly what it's for. Within a single session, the context window handles continuity perfectly. The problem is only between sessions, when the context window resets. That's where persistent memory takes over.

How much of the context window should memory files use?

Aim for 5-10% (5K-20K tokens). This leaves 90%+ for actual conversation. If your memory files are eating more than 20% of the context window, they need trimming — probably by moving old daily notes to archive and keeping only recent ones loaded.

What about models that claim "infinite context"?

No model has infinite context in practice. Even streaming/retrieval approaches have latency and accuracy tradeoffs. The structured memory approach works regardless of context size because it's about organization, not volume. Well-organized 5K tokens beats unorganized 500K tokens every time.

Context Window vs Memory: They're Not the Same Thing