Lyellr88/marm-memory

Local-first 3-in-1 AI memory layer & MCP server for Claude Code, Codex, Grok, Gemini, VS Code and Cursor. Fuses session history, codebase indexing & concept graphs in SQLite. Enables zero-cloud, privacy-first context & instant recall for multi-agent swarms.

What it solves

MARM Memory is a persistent, local-first memory layer designed for AI agents to prevent "agent amnesia," conversational drift, and context pollution. It allows multiple agents (such as Claude Code, Cursor, or VS Code agents) to share a unified, long-term memory across different sessions and projects, ensuring that decisions, code symbols, and project history are not lost when a chat is cleared.

How it works

MARM fuses three distinct context layers into a single runtime using an embedded SQLite database with WAL mode for high concurrency:

  • Core Memory: Handles episodic memory, session logs, and notebooks using a hybrid of local vector embeddings (Jina v2 Small) and full-text search (FTS5).
  • Code Graph: Indexes repositories to provide instant symbol lookup, call tracing, and architecture overviews via static analysis.
  • Concept Graph: Extracts entities and typed relationships from stored history, linking high-level developer decisions directly to specific code symbols.

It operates as a Model Context Protocol (MCP) server, supporting both HTTP and STDIO transports, and includes a local web-based console for visualizing and debugging memory.

Who it’s for

It is primarily built for developers and researchers using MCP-compatible AI coding agents who need a persistent, shared memory state across multiple tools and long-running multi-agent projects.

Highlights

  • Hybrid Retrieval: Combines BM25 exact matching with semantic re-ranking to maintain fast recall speeds even as memory grows.
  • Multi-Agent Support: Uses a serialized write queue to prevent contention when multiple agents are writing to memory simultaneously.
  • Local-First: No external vector database is required; it uses embedded SQLite for storage.
  • Integrated Tooling: Bundles 14 MCP tools covering memory, code indexing, and concept extraction.
  • Visual Telemetry: Includes a built-in local web app (MARM Console) to browse and manage stored knowledge.