If you've ever restarted a Claude or Cursor session and watched your AI relearn everything from scratch, this is the fix. It captures lifecycle hooks from your coding assistant (prompt submits, tool usage, context compaction warnings), runs a local SLM offline to synthesize what happened, and exposes recall, search, recent, and graph_query tools via MCP. The next session picks up with a hot handoff YAML if you're coming back within 30 minutes, or a cold pointer to semantic retrieval if not. Works across Claude Code, Gemini CLI, VS Code Copilot, Codex, and Cursor. Everything stays project-local in .engram-cc/ with no external API calls for its own memory operations.
claude mcp add --transport stdio io.github.morous-dev-engram-cc uvx engram-cc