The report noted that as the user continues to add more messages to the conversation, the oldest details gradually begin to fall out of the model’s available memory, leading to some information being lost or ignored while continuing the conversation.
According to the report, some AI applications rely on internal mechanisms to compress long conversations and automatically recapitulate them when they approach the maximum amount of available context.
For example, Claude’s model might pause the conversation to reorganize and summarize the information. Despite these solutions, the user cannot always know when the system will decide to compress the conversation or delete some old details, and common applications do not often provide a clear option that allows the user to perform this process manually.
The report suggests using a simple approach before moving on to a new conversation, by asking the AI to prepare a “delivery summary” that contains the most important elements of the previous discussion.
It is recommended to use a request that includes specifying the primary goal of the conversation, the decisions that were made, the information that the model might err in inferring if it started from scratch, in addition to the next step required to complete the work.
The report concludes that this method is more effective than simply asking for a general summary of the conversation, because it focuses on the substantive information and maintains the required context without losing important details.
