Context window = a finite, shared resource. Everything in a session competes for the same space: system instructions, conversation history, any retrieved or uploaded content, and the model's own output. It's not just "how much text can Claude read" — it's a budget that gets spent across the whole interaction.
Signs of context strain:
- Degraded recall of earlier parts of the conversation
- Repetition or circling back to points already covered
- Drift — responses gradually losing thread of the original task
Mitigations:
- Start a fresh session for a genuinely new task rather than stretching one thread indefinitely
- Summarize and carry forward only what's load-bearing, rather than the full history
- Use Projects to hold durable, recurring context outside the conversation itself, so it doesn't have to be re-established every session
Memory vs. Project knowledge — a distinction worth being precise about:
- Memory features are what can persist information across separate sessions automatically, without you re-supplying it.
- Project knowledge is context you deliberately attach to a Project, available to every conversation within it, but scoped to that Project rather than following you everywhere. They solve a similar problem (not repeating yourself) but at different scopes — one is automatic and cross-session, the other is deliberate and Project-scoped.