Conversational assistants need to carry useful facts from earlier chats into later ones, but summaries can leave out details. Their memory systems must decide what to keep and what to bring back for each question.
The framework repeatedly edits and tests complete computer programs to build memory systems from simple starting points. These systems store past conversations, retrieve relevant material, and prepare it for answering.
The evolved systems preserve original chats as searchable evidence and organize retrieved information according to the question. On one evaluation, a system scored up to 12 percentage points higher than hand-designed alternatives while using 35% of their running cost. This suggests that preserving source material and tailoring how evidence is presented can improve assistant memory.