Memory right now is flat markdown files (MEMORY.md , SOUL.md) that Claude reads at session start. Honestly it works better than you'd expect. Claude's context window is big enough that a few hundred lines of memory covers most personal assistant use cases. I'd rather keep it as plain text you can read and edit yourself than add a vector DB dependency for marginal recall gains.
OpenClaw doesn't support Claude Code's native OAuth, plus it runs its own daemon. OpenPaw goes the other direction: it writes config files that
Claude Code already understands (CLAUDE.md, permissions, hooks) and exits so claude code can function as an actual personal assistant with all the skills needed. No daemon, no extra process.
https://github.com/daxaur/openpaw