Context Compression

Compression is needed not only for length but also for the quality of thinking: generalized knowledge is more convenient to use than raw knowledge. There are two reasons – a limit on window/cost and context degradation (it fits, but isn’t found). Strategies: context-aware compression (taking into account the current request) is more effective than blind summarization; adaptive window (compress when the threshold is ~80%, in batches); preserving identifiers literally. Isolation is better than compression: a sub-agent performs the dirty work in its own context and returns only the summary to the main agent.

Related: [Agent State String], KV Cache, [Multi-agent interaction without shared context]