Claude Code introduces idle compression: automatically saves tokens after 54 minutes of inactivity, though it may result in lost context details.

Beating AI Express (Dongcha) News: Anthropic has added Idle Compaction to Claude Code, a feature designed specifically to reduce token consumption after long session interruptions. Previously, when users step away from their computers for a period and return to their conversation after prompt caching expires, the model has to reprocess the large volume of prior content. Now, Claude Code automatically compresses long conversations into summaries before the cache expires. When users come back, the model only needs to process the summary instead of the full conversation history, cutting the cost of subsequent API calls. This feature is currently available for the terminal version, with default settings requiring sessions to use 1 hour of prompt caching, a context window of at least 200,000 tokens, and the computer to remain awake. Compaction triggers after approximately 54 minutes of continuous inactivity. However, the cost savings come with tradeoffs: after compaction, Claude can only operate based on the summary, potentially forgetting specific prior instructions, technical details, and decision-making processes. As early as early October, developers complained on GitHub that they were running around 20 long sessions simultaneously, and 7 of them were automatically compacted overnight, forcing them to re-add key context. Anthropic later added an off switch: users can set `"idleCompaction": false` in their `settings.json` file, and also adjust the default 200,000-token trigger threshold.