The problem
Agents forget everything between sessions. Decisions made last month in one project, the reason a library was dropped, the exact thing a client asked for: all of it sits in session transcripts nobody searches.
What I built
memex collects my Claude Code and Codex CLI sessions from every project. A Claude stop hook pushes each session the moment it ends, and an hourly sweep catches the rest. It indexes them in SQLite with full-text search. Each project’s own notes (context, decisions, backlog) stay in that project’s repository; memex indexes them rather than copying them.
There is a web dashboard for me, and an MCP endpoint over HTTP so agents can ask “what did we decide about X” before they ask me. “Active hours” are counted honestly in 30-minute windows of real activity, not wall-clock session length.
It’s private, behind a sign-in, and deployed with Docker through my own deploy tooling.
