Your agent forgets. The workshop remembers.
claude-mem watches the trail of work, distills it into useful notes, and gives the next session a map instead of a mountain of transcript.
From activity to orientation
Nothing mystical happens. claude-mem separates doing the work from writing the memory, then makes recall deliberately cheap.
Observe
Lifecycle hooks notice prompts, tool calls, and session endings.
Compress
A second model turns noisy activity into structured observations.
Index
SQLite stores the memory and FTS5 makes its language searchable.
Recall
The next agent scans a small map, then fetches only useful detail.
A large context window is not the same as clear attention.
A model can receive a great deal of text and still lose the thread. That loss of focus is called context rot. claude-mem protects attention by showing a small index before it spends tokens on full details.
Learn progressive recallLearn it as three small stories
Start anywhere. Each lesson defines the technical language at the moment it becomes useful.
~/.claude-memLocal first is an architectural choice.
The memory database, index, settings, and logs live on your machine. A chosen AI provider processes material for compression. Optional cloud sync can carry the finished memory between machines, but local capture keeps working on its own.