Persistent conversational memory server for MCP-based AI workflows
Waggle-mcp, developed by Abhigyan Shekhar, is an MCP server that provides an external memory layer for AI assistants, recording project reasoning, decisions and state over time. It captures conversational memory, extracts facts into typed graph nodes, and stores them in a local knowledge graph so agents retain context across sessions and applications. The tool offers local sentence-transformer embeddings, temporal queries, and a built-in diagnostic, making it suitable for developers and power users of coding assistants who need persistent project-wide context and privacy-preserving storage.
What tasks can you actually use it for?
The app acts as a conversational memory engine for MCP hosts, recording not only messages but project reasoning, decisions, and changes so agents preserve context after context-window resets. It stores information as typed graph nodes and supports temporal queries to retrieve project state at a given time. Cross-application context sharing enables different MCP clients to access the same memory, which suits multi-tool development flows where continuity across separate chat sessions matters.
What inputs and setup does it require?
It requires Python 3.11 or higher and an MCP host environment such as Claude Desktop, Cursor, Codex, or Antigravity. The project runs on macOS, Linux, and Windows when UTF-8 support is enabled. Installation paths include pipx using the command 'pipx install waggle-mcp' followed by a setup command that auto-configures supported MCP clients. A built-in diagnostic, 'waggle-mcp doctor', checks installation health during initial setup.
Is its memory architecture dependable and suitable for privacy-sensitive projects?
Graph-backed reasoning underpins memory consistency by recording typed nodes and enabling the system to represent contradictions and evolving project logic rather than relying only on vector similarity. The app uses local sentence-transformer embeddings by default, so no external API key is required for vectorization. Local storage uses SQLite primarily, with an optional Neo4j backend available for more complex graph database needs, supporting a local-first privacy model.
Practical choice for developers needing continuous project memory
The tool is a practical option for developers and power users who need persistent project memory across multiple MCP clients. Its local-first design supports privacy-conscious workflows, but adoption depends on integrating an MCP host and meeting the Python 3.11 requirement. For teams that accept local hosting and occasional configuration, it reduces repeated context transfer and helps maintain continuity during multi-session development.
