Shodh Memory

What varun29ankuS/shodh-memory MCP server does

varun29ankuS/shodh-memory MCP server adds persistent memory and task management to MCP clients. Agents can save decisions, preferences, project details, and other context with remember, then retrieve it through semantic, associative, temporal, hybrid, spatial, mission, or action-outcome recall modes. Tag searches, recent-context summaries, memory statistics, and memory listing support inspection and targeted retrieval.

The server also exposes reminders and a GTD-oriented todo system. Agents can create projects and tasks, assign priorities, contexts, due dates, and subtasks, then search, reorder, complete, update, archive, or delete them. Recurring todos create their next occurrence when completed.

How it works

Storage and retrieval run locally rather than sending content to an LLM or a hosted database. The system uses on-device MiniLM embeddings through ONNX Runtime for semantic search. It can extract entities with a local GLiNER model, fall back to rules, and build typed graph relationships from text. Graph traversal supports related-memory retrieval and causal tracing.

Memory is organized into working, session, and long-term tiers. Frequently used memories receive stronger activation, while unused information decays. Recalling one item can activate related items, and repeated connection use can preserve those relationships. Content-hash deduplication prevents identical memories from being stored twice.

The proactive_context tool is intended to run first for every user message. It accepts the message as context, retrieves relevant prior information, and automatically ingests the context for future sessions. Token tools report session usage and can reset the counter after a new or compressed conversation.

Setup and configuration

For an MCP client, install the published npm wrapper with:

npx -y @shodh/memory-mcp

A Claude Desktop or Cursor configuration can run that command with -y and @shodh/memory-mcp as its arguments. Local operation needs no API key; the local setup generates one automatically. Remote deployments can pass a SHODH_API_KEY value to authenticate access.

The project also documents a unified shodh CLI, Python installation, Rust usage, and Docker deployment. The standalone server can listen on port 3030 in the documented Docker example, with a mounted volume used for data persistence. The MCP wrapper is the relevant installation path when an MCP client launches the server locally.

Tools and capabilities

The available capabilities include:

• Store, recall, tag-search, summarize, list, inspect, and delete memories.
• Check vector-index health and re-index orphaned memories.
• Create, list, verify, purge, and restore local backups.
• Review consolidation reports covering strengthening, decay, graph edges, extraction, and maintenance.
• Track token usage and reset a token session.
• Set, list, and dismiss time- or context-triggered reminders.
• Manage todos and projects, including priorities, due dates, subtasks, ordering, completion, and archiving.

Limitations and notes

varun29ankuS/shodh-memory MCP server is designed for local, offline use, but the first run may download the configured local AI model. Its memory intelligence is algorithmic and does not use an LLM in the storage or recall path. Remote deployments require the API-key configuration described in the README, while local clients do not.

Restoring a backup replaces the current user's data, and the documentation recommends restarting the server afterward. The verify_index and repair_index tools are available for diagnosing and repairing memories that exist in storage but are missing from the searchable vector index. The project describes edge-device and robotics support, including Zenoh and ROS2 readiness, but those deployment paths are separate from the basic MCP client setup.