MARM-Systems

MARM MCP provides a local-first memory infrastructure for Claude, Gemini, Qwen, and other MCP clients, enabling AI agents to retain decisions, context, and project history across sessions instead of forgetting after each conversation. It operates as a universal MCP Server supporting HTTP, STDIO, and WebSocket transports, with a memory model built on sessions, structured logs, notebooks, summaries, and semantic memories. The system handles session routing, hybrid recall with semantic re-ranking, serialized writes, and rate-limit presets to support everything from solo developers to multi-agent swarms. MARM uses SQLite with WAL mode, connection pooling, and an intelligent write queue to scale safely without losing data integrity. Developers benefit from persistent project history, shared context across agents, and intelligent recall that stays useful as memory grows.

[{"step":"Install and configure MARM-Systems","action":"Download the Universal MCP Server from [official repository]. Run the server with your preferred protocol (e.g., `./marm-server --protocol websocket --port 8080`). Ensure the MARM protocol is initialized with your desired reasoning structure (e.g., `marm init --reasoning hed`).","tip":"Use Docker for cross-platform consistency: `docker run -p 8080:8080 marm-systems/mcp-server:latest --protocol websocket`"},{"step":"Define memory policies and context schema","action":"Create a memory configuration file (e.g., `memory.yaml`) specifying retention periods, priority flags, and schema for your data types. Load it with `marm config --file memory.yaml`. For multi-agent systems, define shared context keys (e.g., `project_objective`, `team_roles`).","tip":"Start with a 7-day retention window for development and extend to 30+ days for production. Use YAML anchors to avoid repetition in schema definitions."},{"step":"Integrate with your AI workflows","action":"Connect your AI agents/tools to the MCP server. For Python agents, use the `marm-client` library: `from marm import Client; client = Client('ws://localhost:8080'); client.store_context(key='current_task', value='Optimize neural network inference')`. For CLI tools, pipe inputs via STDIO: `echo '{\"task\":\"debug GPU memory leak\"}' | ./marm-server --protocol stdio`.","tip":"Add a `context_id` parameter to all operations to maintain traceability across sessions. Example: `client.query(task='debug', context_id='session_20240610')`."},{"step":"Validate persistence and coordination","action":"Test memory persistence by restarting the server and querying for previously stored context. For multi-agent systems, verify context sharing by having one agent store data and another retrieve it. Use the `marm audit` command to review memory operations.","tip":"Simulate agent handoffs by running separate instances of the MCP server on different ports (e.g., 8080 for Agent A, 8081 for Agent B) and configure them to sync via a shared memory backend (e.g., Redis)."},{"step":"Monitor and refine","action":"Use the MARM dashboard (available at `http://localhost:8080/dashboard`) to monitor memory usage, reasoning traces, and agent interactions. Adjust retention policies or schema based on observed patterns. For debugging, enable verbose logging with `marm server --log-level debug`.","tip":"Set up alerts for memory limits (e.g., 80% capacity) to prevent data loss. Use the `marm export` command to back up critical context before major updates."}]