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, supports multi-agent swarms.

What it solves

AI agents typically lose context between chat sessions, forcing them to start from scratch each time. MARM Memory provides a persistent, local-first memory server that allows agents to store and recall decisions, research, and project history across different sessions and even across different AI models (e.g., switching from Claude to Gemini).

How it works

MARM operates as a Model Context Protocol (MCP) server that provides a suite of 14 tools to the connected agent. It uses a hybrid retrieval system combining Full-Text Search (FTS5 BM25) and semantic re-ranking to ensure fast and accurate recall as the memory grows. The system is organized into three layers:

  • Core Memory: Stores conversations, notes, and summaries.
  • Code Graph: Indexes repositories to map symbols and code paths, allowing agents to understand project structure without rereading the entire codebase.
  • Concept Graph: Extracts entities and relationships from memories, linking them back to the code graph.

It uses an embedded SQLite database with WAL mode and a serialized write queue to handle concurrency, and can be deployed via pip or Docker using various profiles (solo, swarm, or trusted).

Who it’s for

Developers and researchers who use MCP-compatible AI agents (like Claude Code, Codex, or Gemini) and want their agents to have a permanent, shared local memory of their projects and decisions.

Highlights

  • MCP-Native: Works with any MCP client via HTTP or STDIO transports.
  • Hybrid Retrieval: Combines keyword search and semantic embeddings for efficient scaling.
  • Code & Concept Graphs: Goes beyond flat text by mapping repository structures and conceptual relationships.
  • Local-First: Uses embedded SQLite for zero-config local setup.
  • Web Console: Includes a built-in UI to browse memories, knowledge graphs, and indexing progress.

Related

  • Project
    doobidoo/mcp-memory-service

    A persistent, shared memory backend for AI agents that provides a knowledge graph and semantic search across sessions and tools via REST API and MCP.

  • Project
    rohitg00/agentmemory

    A persistent memory server for AI coding agents that automatically captures and retrieves project context across sessions using hybrid search.

  • Project
    agentic-box/memora

    An MCP memory layer for AI agents that provides persistent, structured storage and semantic retrieval of facts, tasks, and documents via a knowledge graph.

  • Project
    Gentleman-Programming/engram

    Engram is a single‑binary Go program that gives AI coding agents a persistent, searchable memory store (SQLite + FTS5). It works with any MCP‑compatible agent (Claude Code, Gemini CLI, VS Code Copilot, Cursor, etc.) via CLI, HTTP API, MCP, or a terminal UI. The tool is zero‑dependency, supports local‑first storage with optional cloud replication, and provides a rich set of memory commands (`mem_save`, `mem_search`, `mem_context`, `mem_session_summary`, …) to let agents remember decisions, bug fixes, and design rationales across sessions and teams.

  • Project
    JuliusBrussee/cavemem

    A cross-agent persistent memory system for coding assistants that uses compressed local storage and MCP to allow AI agents to remember information across sessions.