Agent Memory Mcp

Markdown files as a shared memory layer across AI tools. Saves rules, decisions, and context with severity levels and enforcement categories, then auto-projects them to AGENTS.md, CLAUDE.md, Cursor MDC files, and Gemini instructions so every tool sees the same constraints. The check_action tool gates destructive operations by matching proposed actions against your rule store and issuing cryptographically signed compliance receipts with 60 second TTLs. Breaking change in v0.12: delete_memory requires a valid receipt. You can grep the memory directory, edit rules in vim, commit to git, and sync across machines. Falls back to regex matching when sampling isn't available, so enforcement works even in clients without LLM-in-the-loop support.