MARM Memory is a local-first MCP memory server that fuses session memory, a code graph, and a concept graph into one embedded SQLite store, exposing 14 tools over STDIO or HTTP. It suits self-hosters who want persistent context shared across multiple AI coding clients without sending memory to the cloud; recall uses an FTS5 exact lane plus semantic reranking, and the project's own benchmarks show recall latency scaling slowly with memory size. Note that it is not an official vendor product, and HTTP exposure, Docker mounts, and optional consolidation/compaction features add setup complexity and real performance tradeoffs, so verify keys, network binding, and data-directory permissions before deploying.
Strongest · Documentation 15/20 Weakest · Reliability 10/20
Static review · not runListed 2026-09-10
Read the FMRS scoring method →Runtime：Python 3.10+ · Docker 20.10.0+
MARM_API_KEY
Bearer key required for Docker or exposed HTTP mode; not needed for local pip/STDIO. Generate via marm-memory key init or docker run --rm lyellr88/marm-mcp-server --generate-key.
MARM_DB_PATH
Path to the SQLite database file, default /app/data/marm_memory.db.
MARM_LOG_LEVEL
Logging level (DEBUG/INFO/WARNING/ERROR), default INFO.
SERVER_HOST
Host address to bind to, default 0.0.0.0; no key needed for localhost-only use.
SERVER_PORT
HTTP port to bind to, default 8001.
RECALL_SCAN_LIMIT
Max embedded memories the semantic fallback scans per query (default 10000); exceeding it sets recall_scan_truncated=true.
MARM_PROJECT
Overrides automatic project detection from the working directory.
MARM_PLATFORM
Overrides automatic platform (client) detection.
GRAPH_ENABLED
Set to false to disable the graph engines; graph tools then return a degraded-mode error.
CONCEPT_AUTO_INDEX
Set to false to stop automatic indexing of new memories into the concept graph (manual builds only).
CONCEPT_INDEX_DEBOUNCE_SECONDS
Debounce interval for background concept indexing, default 30 seconds.
CONCEPT_INDEX_BATCH_SIZE
Memories processed per concept-indexing batch, default 20.
CONCEPT_BUILD_ROW_CAP
Page size for full concept builds, default 500.
CONSOLIDATION_ENABLED
Set to 1 to enable write-time dedup/merge (~9x write cost), off by default.
CONSOLIDATION_THRESHOLD
Cosine similarity threshold for semantic merge during consolidation.
COMPACTION_ENABLED
Set to 1 to enable detection of memory compaction candidates.
COMPACTION_STAGING_TTL_HOURS
Expiry time in hours for staged compaction summaries.
MAX_QUEUE_SIZE
Upper bound on the internal serialized write queue.
FTS_CANDIDATE_LIMIT
Cap on FTS pre-filter candidates in hybrid recall, default 200.
GRAPH_AUTO_INDEX
Environment toggle for repo auto re-indexing; the persistent switch takes precedence.
{
"mcpServers": {
"marm-memory-local": {
"type": "http",
"url": "http://localhost:8001/mcp"
},
"marm-memory-docker": {
"type": "http",
"url": "http://localhost:8001/mcp",
"headers": {
"Authorization": "Bearer ${env:MARM_API_KEY}"
}
}
}
}
Shown for Claude Code. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.
{
"servers": {
"marm-memory-local": {
"type": "http",
"url": "http://localhost:8001/mcp"
},
"marm-memory-docker": {
"type": "http",
"url": "http://localhost:8001/mcp",
"headers": {
"Authorization": "Bearer ${env:MARM_API_KEY}"
}
}
}
}
Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).
claude mcp add --transport http marm-memory-local http://localhost:8001/mcp
claude mcp add --transport http marm-memory-docker http://localhost:8001/mcp --header 'Authorization: Bearer ${env:MARM_API_KEY}'
Run it in a terminal; replace any <…> placeholders with your own values first.
After marm-memory start, run curl http://localhost:8001/health to confirm the service is healthy; once connected, the client should list the 14 tools such as marm_log_entry and marm_smart_recall, and an agent can store a test memory with marm_log_entry then retrieve it with marm_smart_recall.
Once connected, you can ask your AI assistant things like:
No matching tools
Listed from the project's documentation, not tested by this site.
MARM Memory is a local-first MCP server that gives AI agents a permanent, private memory across Claude Code, Codex, Grok, Gemini, VS Code, and Cursor. It fuses session history, repository code indexing, and a concept graph into an embedded SQLite store, enabling zero-cloud, privacy-first context retention and instant recall, including for multi-agent swarms. The server exposes 14 MCP tools: 7 core memory tools (logging, hybrid recall, summaries, notebooks, log viewing, deletion, and compaction) plus 5 code-graph tools (repo indexing, symbol lookup, call tracing, architecture overview, change-impact analysis) and 2 concept-graph tools (entity/relationship build and query). Recall uses a hybrid pipeline of an FTS5 BM25 exact lane plus semantic reranking, with long memories chunked into overlapping rows and collapsed back to the parent memory. Both STDIO and HTTP transports are supported, with managed swarm, swarm-max, and trusted rate-limit profiles for shared deployments.
Upstash's official server providing up-to-date third-party library docs for AI coding assistants
AI-powered NPM package analysis MCP server
An independent control layer that gives AI coding agents budgets, brakes, and receipts
GitHub's official MCP server for managing repos, issues, PRs, and workflows via natural language
Source revision 97871d57ce7a Data synced 2026-10-19 Read the FMRS scoring method