Context window: AI’s short-term memory

1. At a Glance

2. Why in the News

3. Background & Evolution

4. Core Static Facts

Item Detail
Definition Maximum amount of text (in tokens) an AI model can consider simultaneously while generating a response [S1]
Token-to-word ratio ~1 token = 0.75 words (English); 1,000 tokens ≈ 750 words [S1]
What the window holds (i) System/behavioural rules, (ii) conversation history, (iii) space reserved for the model's next answer — all sharing one budget [S1]
Overflow behaviour If conversation exceeds the limit, the model may start deleting/truncating the oldest parts of the exchange [S1]
Compute scaling Doubling context window length increases computational power required by roughly 4× (non-linear/quadratic-type scaling in attention mechanisms) [S1]
Known failure mode "Lost in the middle" — models struggle to retrieve information buried in the centre of a long context even if technically within the window [S1][S3]
Example models cited GPT-5, Claude (Hindu article); Gemini 1.5 (million-token context) [S1][S2]

5. Multi-Dimensional Analysis

6. Recent Developments (last 12-18 months)

7. Prelims Hooks

8. Mains Relevance

9. Related Topics to Study Next

10. Common Errors / Trap Areas

11. Sources