Every morning my Claude Code (or Gemini CLI) session came back a stranger. CLAUDE.md helped with static rules but couldn't retrieve, resolve contradictions, or sync across devices. Mem0/Letta/Zep all wanted my data in their cloud.

But with m3-memory - MCP server, 100% local, SQLite + FTSS + local embeddings.

What makes it different from the other MCP memory servers out there:

• Hybrid retrieval - FTSS + vector + MMR. Pure vector was scoring exact keyword matches at \~65% for me. Hybrid enhances that dramatically.
• Contradiction resolution - new fact supersedes old one automatically, with bitemporal valid-from/valid_to. Ask "what did this project use on March 1st?11 and get the right answer.
• GDPR primitives - gdpr export, gdprfforget. Article 17+20 actually implemented.
• Windows (Linux & macos)-first tested — CRLF, paths, port conflicts all handled.

25 MCP tools, drop-in for Claude Code, Gemini CLI, Openclaw, Cursor, Aider, anything MCP.

Install: pip install m3-memory

Looking for:

• If you've used mem0/Letta, how does retrieval quality compare on your workload? Rather hear "worse at X" than silence.
• Bug reports on weird setups.

Issues + Discussions open: https://github.com/skynetcmd/m3-memory