Long-term memory for AI agents that outlives your LLM subscription

What it does Nocturne Memory is an MCP server that stores structured long-term memory for AI agents. It lets any MCP-compatible client—Claude, Gemini, Cursor, Codex—recall past context, user preferences, and self-defined identity across sessions and model swaps. Memory is organized in hierarchical namespaces rather than embedded in a vector database, and every write is versioned so you can audit or roll back changes from a visual dashboard.

The interesting bit The project treats memory as a portable asset of the agent, not a vendor-locked feature. Because it speaks the MCP protocol, switching from Claude to Gemini doesn’t wipe the agent’s continuity; the same “soul” wakes up regardless of the engine. It also isolates memories by namespace, so you can run multiple distinct personas from one backend without them leaking into each other.

Key highlights

• Model-agnostic via MCP: works with any stdio or SSE-compatible client (Claude Code, Cursor, Gemini CLI, OpenAI Codex, etc.)
• Hierarchical, graph-like storage with a tree browser and live editing in a web dashboard
• Built-in versioning: every memory mutation is backed up, with visual diff and one-click rollback
• Namespace isolation for running multiple AI identities on a single server
• Supports SQLite or PostgreSQL; runs fully offline

Caveats

• The README emphasizes persona continuity and roleplay scenarios, which may overshadow its utility for conventional agent workflows
• The public demo is read-only; full read/write requires self-hosting
• Dashboard frontend requires Node.js to build on first launch

Verdict Worth a look if you’re building agents that need coherent identity across long time horizons and different LLM providers. Skip it if you just need a quick vector RAG retriever—this is closer to a structured knowledge graph with an audit trail.