Building AI Agents in Kotlin – Part 5: Teaching Agents to Forget | The JetBrains AI Blog

In the fifth installment of a series on building AI agents in Kotlin, the focus is on teaching agents to manage their memory by implementing a strategy of smart compression to prevent context overflow. The series has been utilizing the GPT-5 Codex model, which performs well on simple tasks but struggles with complex tasks due to context window limitations. A new approach, singleRunStrategyWithHistoryCompression, is introduced, which allows agents to compress their history intelligently, retaining crucial context while discarding verbose details. This compression is triggered when history size exceeds set thresholds, ensuring that agents can handle longer tasks without crashing due to context limits. Users can customize the compression strategy by defining specific concepts that the agent must remember, thus maintaining the necessary information to continue tasks effectively. This enhancement enables agents to perform tasks efficiently within token budgets and lays the groundwork for further exploration into agent planning and reasoning. The complete implementation is available on GitHub for those interested in further development.