Mémoire résumée ou mémoire verbatim
Two ways of giving an agent the memory of your past sessions, and they do not lose the same thing. **Summarised** memory runs each session through a model that keeps the gist, then feeds that digest back into later sessions: that is the claude-mem approach, which captures Claude Code activity through hooks, compresses it with an AI and restores it at startup. **Verbatim** memory keeps the original text as it is and only chooses at search time: that is the MemPalace approach, which stores conversations word for word and finds them again through semantic search, with no summarising or rewording. The first reads back fast but has already decided what mattered; the second keeps the whole reasoning, at the price of a search you have to phrase well. Both tools sit one after the other in a list of recommended plugins by Emilien Corbineau, and the list never says they answer the same need in two opposite ways.