An AI agent’s memory is a mechanism that allows it to reference information gained from previous conversations or tasks in subsequent work. However, if errors are stored in the memory, those errors may be carried forward. Anthropic is testing a system that separates working notes from original records, enabling later comparison and correction. By structuring the system to allow users to trace back from summaries to the underlying source records, errors introduced in the notes can be corrected. Additionally, during contract evaluations, multiple agents rewriting the same table inadvertently overwrote and lost critical terms. By assigning separate pages and responsible agents for each contract and preserving clauses exactly as they appear, information loss was significantly reduced. The mechanism for retrieving memory is also crucial. If users can navigate from a brief table of contents to the specific details they need, there is no need to reread everything each time. However, well-organized memory is not always superior to direct search—its effectiveness varies depending on the task. This explanation highlights that designing memory systems requires careful consideration not only of memory capacity, but also of error correction, assignment of recording responsibilities, and efficient pathways to retrieve necessary information.