marm-memory

Name: marm-memory
Author: Ryan Lyell

Local SQLite memory for coding agents, with recall, a code graph and a concept graph

By Ryan Lyell (individual) Python Local · stdio, Streamable HTTP Knowledge and memory Developer tools

GitHub PyPI Docker Hub Website

54 popularity score ★ 434 ↓ 4.1k/mo

Connect your client

Local (once it is running)

Add in one click

Add to Cursor Add to VS Code Add to LM Studio Add to Goose

Opens the app, which asks before it adds the server.

Or add it by command or config

Claude CodeTerminal

claude mcp add --transport http marm-memory http://localhost:8001/mcp

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "marm-memory": {
      "url": "http://localhost:8001/mcp"
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "marm-memory": {
      "type": "http",
      "url": "http://localhost:8001/mcp"
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.marm-memory]
url = "http://localhost:8001/mcp"

Gemini CLITerminal

gemini mcp add --scope user --transport http marm-memory http://localhost:8001/mcp

Local (PyPI)

Add in one click

Add to Cursor Add to VS Code Add to LM Studio

Opens the app, which asks before it adds the server.

Or add it by command or config

FirstTerminal

pip install marm-mcp-server

Claude CodeTerminal

claude mcp add marm-memory -- marm-mcp-stdio

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "marm-memory": {
      "command": "marm-mcp-stdio",
      "args": []
    }
  }
}

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "marm-memory": {
      "command": "marm-mcp-stdio",
      "args": []
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "marm-memory": {
      "type": "stdio",
      "command": "marm-mcp-stdio",
      "args": []
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.marm-memory]
command = "marm-mcp-stdio"
args = []

Gemini CLITerminal

gemini mcp add --scope user marm-memory marm-mcp-stdio

Local (Docker)

Add in one click

Add to Cursor Add to VS Code Add to LM Studio Add to Goose

Opens the app, which asks before it adds the server. Then replace each <NAME> in its settings with your value.

Or add it by command or config

Claude CodeTerminal

claude mcp add marm-memory -- docker run --rm -i -v '<ABSOLUTE_PATH>:/home/marm/.marm' --entrypoint python lyellr88/marm-mcp-server:latest -m marm_mcp_server.server_stdio

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "marm-memory": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "-v",
        "<ABSOLUTE_PATH>:/home/marm/.marm",
        "--entrypoint",
        "python",
        "lyellr88/marm-mcp-server:latest",
        "-m",
        "marm_mcp_server.server_stdio"
      ]
    }
  }
}

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "marm-memory": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "-v",
        "<ABSOLUTE_PATH>:/home/marm/.marm",
        "--entrypoint",
        "python",
        "lyellr88/marm-mcp-server:latest",
        "-m",
        "marm_mcp_server.server_stdio"
      ]
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "marm-memory": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "-v",
        "<ABSOLUTE_PATH>:/home/marm/.marm",
        "--entrypoint",
        "python",
        "lyellr88/marm-mcp-server:latest",
        "-m",
        "marm_mcp_server.server_stdio"
      ]
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.marm-memory]
command = "docker"
args = ["run", "--rm", "-i", "-v", "<ABSOLUTE_PATH>:/home/marm/.marm", "--entrypoint", "python", "lyellr88/marm-mcp-server:latest", "-m", "marm_mcp_server.server_stdio"]

Gemini CLITerminal

gemini mcp add --scope user marm-memory docker -- run --rm -i -v '<ABSOLUTE_PATH>:/home/marm/.marm' --entrypoint python lyellr88/marm-mcp-server:latest -m marm_mcp_server.server_stdio

Settings 10

`SERVER_HOST`	Optional: HTTP bind address, 127.0.0.1 by default; 0.0.0.0 exposes it and makes MARM_API_KEY mandatory.
`MARM_API_KEY`	Optional: Bearer key for HTTP when network-exposed or in Docker HTTP; not used over STDIO.
`SERVER_PORT`	Optional: HTTP port, 8001 by default.
`MARM_DB_PATH`	Optional: Memory database file, ~/.marm/marm_memory.db by default.
`MARM_RATE_LIMIT_RPM`	Optional: HTTP requests per minute per IP, 80 by default; 0 turns limiting off.
`GRAPH_ENABLED`	Optional: Set to false to switch off the six code-graph tools.
`SEMANTIC_SEARCH_ENABLED`	Optional: Set to 0 to run without the embedding model (keyword recall only).
`CONSOLIDATION_ENABLED`	Optional: Set to 1 to merge near-duplicate memories at write time (off by default).
`MARM_LLM_ENABLED`	Optional: Turn on optional generation through a local model server (off by default).
`MARM_LLM_URL`	Optional: Local OpenAI-compatible endpoint for generation, http://127.0.0.1:18080 by default.

In a container, a setting goes in with -e NAME=value after docker run.

What it does

marm-memory (published on PyPI as marm-mcp-server) keeps a store of notes, session logs and summaries on your own machine so that an assistant can pick up context in a later chat, or in a different client. Everything sits in SQLite files under ~/.marm. Recall mixes a keyword pre-filter with re-ranking by a local sentence-embedding model, and the same store can be shared by several agents when the server runs over HTTP.

Beyond plain memory it bundles two graphs. A code graph indexes a local repository (through the separate codebase-memory-mcp engine) so an agent can look up symbols, trace calls, see an architecture overview and estimate what a git diff touches; once a repo is indexed it is re-indexed as files change. A concept graph pulls people, decisions, errors and ideas out of stored memories and links them back to code. Tools also cover deleting entries, a session notebook, reviewed clean-up of near-duplicate memories, and turning a pasted conversation into proposed memories.

It runs over STDIO with no key, or as a local HTTP service at http://localhost:8001/mcp that several clients can share; the pip package also ships a CLI (marm-memory) and a browser console on port 8002. At v2.60.1 the STDIO entry point registers 16 tools: eight memory tools defined in server_stdio.py and eight graph tools registered from a separate module. In that file marm_smart_recall returns 5 results by default, searches only the current session ("default") unless search_all is set, and returns about 200 characters per result unless detail is raised to 2 or 3.

What it needs

Runtime	Python 3.10+orDocker

Python 3.10 or newer with pip (pip install marm-mcp-server), or Docker for the container. Runs on Windows, macOS and Linux; the docs ask for about 100 MB of disk and 512 MB of RAM. The first run downloads the embedding model, and on pip installs the first graph tool call downloads the codebase-memory engine binary, so both need internet access once. HTTP mode is started separately (marm-memory start) and clients then connect to http://localhost:8001/mcp. Optional text generation needs a local OpenAI-compatible model server.

Limits

Recall returns 5 results from the current session by default; set limit, or search_all to look across sessions. If the graph engine cannot start (no network for its first download, disk full) or GRAPH_ENABLED=false is set, the code-graph tools return an error while the memory tools keep working. In Docker the graph tools see only paths mounted into the container, and a mount cannot be added to a running container. Memories are written only when the agent calls a tool; nothing is captured on its own. The makers' own retrieval benchmark (LoCoMo) puts multi-hop questions at 44.9%. Write-time de-duplication and compaction are off by default.

Who it suits

Suits developers who switch between several coding assistants and want one private, local memory they can all reach, especially when they also want repository structure on hand without re-reading the code. STDIO suits one agent on one machine; HTTP suits several agents sharing a store. It is maintained by one developer, not a vendor. Anyone wanting hosted or synced memory across machines should look elsewhere: it keeps everything in local files and has no cloud sync.

Tools 16

marm_smart_recall: Search stored memories by keyword and meaning, with related graph context when it exists
marm_log_entry: Add an entry to a session log; it also becomes searchable memory
marm_log_show: List sessions and their log entries
marm_delete: Remove a log session, log entry or notebook entry
marm_summary: Produce a summary of a session ready to paste elsewhere
marm_notebook: Keep session notes and save them as lasting, graph-linked documents
marm_compaction: Find, review and apply merges of near-duplicate memories
marm_distill: Turn a raw conversation into memory proposals checked against the store
marm_graph_index: Index a repository into the code graph, list projects, toggle re-indexing
marm_code_lookup: Find symbols, text or a symbol's source in indexed code
marm_code_context: Gather ranked symbols, their source and related memories for one task
marm_graph_trace: Follow call paths and data flow from a function
marm_graph_architecture: Show modules, node and edge counts and schema of an indexed repo
marm_graph_impact: Estimate which symbols a git diff affects and how risky it is
marm_concept_build: Rebuild the concept graph or index older memories into it
marm_concept_recall: Query entities, relationships and linked code symbols in the concept graph

Popularity

54 popularity score, on our list (the bar is 36)

★ 434 ↓ 4.1k/mo

GitHub stars, whole repository	434	24.1 points
Stars in 30 days	+15	6 points
PyPI downloads, last month	4,089	24.1 points
Docker Hub pulls	58,855

It lives in the folder marm-mcp-server of lyellr88/marm-memory, which holds more than this server. Its stars are the whole repository's, so in the score they never count for more than its own downloads.

Measured 2026-11-08. How the score works.

Where it is published

Code	lyellr88/marm-memory/marm-mcp-server ★ 434, the whole repository
PyPI	marm-mcp-server 2.60.1, 2026-10-12, the one the command runs
Docker Hub	lyellr88/marm-mcp-server

What it can reach

Runs	On your machine
Reaches	Files on your machine, A database, Your code, The internet
Changes things	Yes: it can write, send, create or delete
Sign-in	None
Licence	Apache-2.0 (PyPI)
Last code change	2026-10-16

Runs on your machine; memories and indexes are SQLite files under ~/.marm, and its docs state there is no cloud sync or telemetry. Tools write and delete memory entries and read source files of repositories you index. Network use: a one-time download of the embedding model and, on pip installs, of the codebase-memory engine binary at the first graph call. STDIO needs no key. HTTP binds to 127.0.0.1 without a key unless SERVER_HOST=0.0.0.0 is set, which makes a MARM_API_KEY bearer token mandatory; Docker HTTP always needs the key. Optional generation calls a local model server and refuses a non-loopback address unless MARM_LLM_ALLOW_REMOTE is set to an exact confirmation string. The bundled console (not an MCP tool) includes a local shell terminal.

Over time

GitHub stars, whole repository 434 +15 (+4 %)

434419

9 Oct 20268 Nov 2026

PyPI downloads a month 4.1k +2.4k (+137 %)

5.2k1.5k

2 Jun 20268 Nov 2026

Before the site measured it, read from pypistats' daily download counts.

Popularity score 54 −1 (−3 %)

5648

9 Oct 20268 Nov 2026

What is rising across the directory

Sources

Page checked 2026-10-14.

Badge for your README

Shows this server's rank in its category, and stars for a server that runs on your machine. It updates every hour.

MarkdownREADME.md

[![marm-memory on EveryMCPServer](https://everymcpserver.com/badge/marm-memory.svg)](https://everymcpserver.com/servers/marm-memory)

Connect your client

Add in one click

Or add it by command or config

Add in one click

Or add it by command or config

Add in one click

Or add it by command or config

Settings 10

What it does

What it needs

Limits

Who it suits

Tools 16

Similar servers

Engram

State Memory MCP

Signet

MegaMemory

projectmem

Claude-Mem