The article examines Claude's memory system, highlighting its distinct differences from ChatGPT's. The author reverse-engineered Claude's architecture and found that it relies on on-demand tools and selective retrieval instead of pre-computed summaries. This approach allows Claude to maintain a rolling window of recent messages while also offering the option to search past conversations when needed. Claude's memory features, however, are only available to paid users, and the author emphasizes that some responses can vary due to the model's non-deterministic nature. In terms of structure, Claude's prompt includes a system prompt with static instructions, user memories that function similarly to long-term memory, and a sophisticated conversation history mechanism. User memories are stored in an XML-like format, containing specific details such as the user's name, preferences, and projects. Claude updates these memories implicitly during conversations or through explicit user commands like "Remember this" or "Delete this from memory." The conversation history is more dynamic, allowing Claude to drop older messages as it reaches a token limit and selectively retrieve relevant past exchanges. The author outlines the methodology used for this analysis, noting that Claude was more cooperative than ChatGPT, making it easier to gather information. The key differences in memory management show that Claude can access detailed historical context when necessary, but it also relies on the model's ability to recognize when that context is relevant. This selective approach contrasts with ChatGPT's always-on summary method, making Claude's memory system more efficient in some respects, but potentially less reliable if the model fails to pull in pertinent information.