how-it-works

Claude-mem captures observations by compressing Read, Edit, and Bash actions into summarized data during an active session. These observations are then processed at session end to create a final memory summary.

Memory injection occurs in subsequent sessions after the first session seeds the memory. The initial session generates and stores the compressed observations, and future sessions automatically receive this injected context.

Claude-mem stores session memory data locally in the ~/.claude-mem directory. This data locality ensures observations remain on your machine unless external providers are explicitly configured for compression.

Yes, claude-mem can send observation data to external providers for compression if explicitly configured. Without this configuration, all memory data remains strictly local in ~/.claude-mem.

The data flow starts with capturing Read, Write, and Bash actions, compressing them into observations, summarizing at session end, and storing locally in ~/.claude-mem. Subsequent sessions then retrieve and inject this stored context automatically.

Claude-mem uses a two-session flow because the first session must seed memory by capturing and compressing actions into observations. Subsequent sessions can then retrieve and auto-inject this context, ensuring accurate historical data.