Self-Compacting Language Model Agents

Apple researchers have introduced a new method for managing the memory of AI agents by compacting long sequences of reasoning and tool use based on their relevance rather than fixed length. By selectively removing stale information from a model's context, this approach prevents critical memory limits from being reached prematurely during complex tasks. This development improves how language models maintain focus during extended operations, potentially allowing agents to complete multi-step assignments more reliably without losing track of earlier instructions.