github.com/mattalldianhr/claude-session-memory
A SessionStart hook scans for unharvested transcripts; a Python pipeline calls a local Gemma 4 model on the homelab to extract structured learnings — debugging solutions, API discoveries, configuration notes, workflow patterns — and writes them to AutoMem, a hybrid graph-plus-vector store.
The same store is exposed back to Claude Code through an MCP bridge, so the next session can ask what previous ones figured out.
About sixteen thousand memories live there now, harvested across every project on this machine.