Lyellr88/marm-memory

pip install marm-mcp-server
marm-memory init --g-claude --g-codex --g-gemini (installs the marm-init skill into agent configs)
Docker: docker run -d --name marm-mcp-server -p 127.0.0.1:8001:8001 -v ~/.marm:/home/marm/.marm lyellr88/marm-mcp-server:latest

Usage

• •Start the server: marm-memory start (HTTP on localhost:8001, keyless on loopback)
• •Connect a client: 'claude mcp add --transport http marm-memory http://localhost:8001/mcp' (Codex: codex mcp add marm-memory --url http://localhost:8001/mcp)
• •Or STDIO: 'claude mcp add --transport stdio marm-memory-stdio marm-mcp-stdio'
• •Just tell the agent in natural language (e.g. 'log this session'); the agent calls the tools automatically

Key features

• ✓14 MCP tools: marm_smart_recall, marm_log_entry, marm_log_show, marm_delete, marm_summary, marm_notebook, marm_compaction + 5 code-graph tools (index/lookup/trace/architecture/impact) + 2 concept-graph tools
• ✓Hybrid recall: an exact FTS5 BM25 lane for config keys/API names plus a filter→rerank semantic lane (jina-embeddings-v2-small-en, 512-dim) with bounded semantic fallback
• ✓Code graph via codebase-memory-mcp (tree-sitter, 158 languages) with git-signature auto-reindexing; concept graph of entities/relationships with code cross-linking
• ✓SQLite WAL + connection pooling + serialized write queue; local-first, everything in ~/.marm/ (no cloud sync, no telemetry)
• ✓HTTP and STDIO transports; swarm/multi-agent presets with rate limits (80/200/600 RPM); web Console

Best for

Developers who want persistent, fully-local cross-session and cross-agent memory for coding agents (Claude, Gemini, Codex, Qwen) with code-structure awareness.

Caveats

• ⚠Requires Python 3.10+, ~100MB storage minimum, 512MB RAM minimum
• ⚠Docker HTTP mode requires an API key (MARM_API_KEY); local pip HTTP is keyless only on loopback (SERVER_HOST=0.0.0.0 forces key mode)
• ⚠Graph engine binary (~269MB) is downloaded on first graph use for pip installs; graph tools degrade if it fails, core memory keeps working
• ⚠Write-time consolidation (~9x write cost) and compaction are opt-in; first semantic search loads the embedding model (slower)
• ⚠Concept-graph extraction is CPU-bound; draining a backlog raises median recall from ~8ms to ~16ms temporarily
• ⚠Distributed multi-node memory is out of scope (one machine, many agents)