Session Memory

The Problem

Running several active codebases at once in Claude Code or Codex, the usual ways of giving an AI coding assistant memory both break down in practice. Re-explain the project's state at the start of every session, and real time is lost to it. Keep one growing memory file instead, and every session loads irrelevant history from projects it isn't even touching.

How It Works

Session Memory takes a narrower approach. Each project gets its own small set of files: a running log of what happened, a log of what actually went wrong — not everything, only genuine failures — and whatever backlog is already kept. Ending a session writes to whichever of those actually changed. Starting one reads only that project's files and gives a short brief: what's done, what's next, what to watch for.

Two Modes

A deliberately small piece of infrastructure: two modes, a handful of markdown files, and clear rules about when to write versus skip. Work across several projects, and an optional shared index tracks which one was touched most recently — including ones with no git history, using a small timestamp file rather than commit dates. Nothing is shared between projects unless that's turned on.