Context Window: Why AI "Forgets"

How Context Windows Work

• Context window sizes vary by model
  • › Claude Opus: ~1M tokens (~750K words)
  • › Claude Sonnet: ~200K tokens (~150K words)
  • › GPT-5.x models: ~128K–200K tokens depending on variant
  • › Gemini 2.5 Pro: ~1M tokens
• When your conversation fills the window, older messages get compressed or dropped — the AI literally can't see them anymore.
• Symptoms of context overflow
  • › AI repeats work it already did
  • › AI 'forgets' decisions you agreed on
  • › AI contradicts its earlier responses
  • › Code quality degrades as the session goes on
• The mental model — context window = short-term memory. CLAUDE.md = long-term memory. Write down what matters.

How to Manage It

1. 1 Prevention strategies
   • › Start fresh for new tasks — don't chain unrelated work in one session
   • › Use subagents for expensive operations — they get their own context
   • › Put conventions in CLAUDE.md — re-loaded automatically, never forgotten
   • › Compact your context when sessions get long — it summarizes and frees space (but loses nuance)
2. 2 Watch your token usage — start a new session when you switch to a different task
3. 3 Move important conventions and decisions into CLAUDE.md so they survive across sessions
4. 4 Compact context when a session gets long but you're not ready to start fresh — it summarizes and frees space