I tried 3 different ways to fix Claude Code's memory problem — here's what actually worked

I lost a full afternoon's work last week. Four hours of debugging a gnarly database migration, maybe 30 back-and-forth messages with Claude Code about the exact schema evolution. Then I closed the terminal, came back after dinner, started a new session — and Claude had no idea what we'd figured out. The CLAUDE.md I'd written was vague. The git commits were there but the why was gone. This is the Claude Code memory problem, and if you code with AI tools regularly, you've hit it. Each session is a blank slate. The context you built up — the decisions, the dead ends, the "wait, we tried that and it failed because..." — evaporates. I spent the last couple weeks trying three different approaches. Here's what happened. The problem in concrete terms Before I get into solutions, let me be specific about what I was trying to solve. There are actually two distinct problems people conflate: Problem A — Intra-session compaction : Claude Code compacts conversation history during a long session to s