Perhaps the trick used by text-based LLMs could be used: when the context window starts filling up, the LLM is asked to summarize the existing data in the conte

Perhaps the trick used by text-based LLMs could be used: when the context window starts filling up, the LLM is asked to summarize the existing data in the context, thus compressing it (lossily..) into smaller space.