Memory MCP

A SQLite-backed MCP server for persistent memory storage, full-text retrieval, and relationship graph traversal.

Overview

Memory MCP provides a local, persistent memory layer for MCP-enabled assistants. It stores SHA-256-addressed memory items in SQLite with FTS5-powered full-text search, a directed relationship graph, BFS recall traversal, and token-budget-aware context retrieval — all accessible over stdio transport with no external dependencies.

Key Features

• 13 MCP tools for CRUD, batch operations, FTS5 search, BFS graph recall, token-budget context retrieval, relationships, and stats.
• Full-text search over content and tags via SQLite FTS5 with importance and type filters.
• Graph recall with BFS traversal, bounded frontier, and MCP progress notifications per hop.
• Token-budget retrieval (retrieve_context) selects memories that fit a caller-specified token budget — no manual pagination needed.
• Strict Zod input validation with typed output envelopes and SHA-256 hash addressing.
• Resource support with internal://instructions (Markdown guide) and memory://memories/{hash} URI template with hash auto-completion.
• stdio transport with clean shutdown handling (SIGINT, SIGTERM) and no HTTP endpoints. |

Requirements

• Node.js >=24.
• SQLite with FTS5 support (verified at startup).
• Any MCP client that supports stdio command servers.

Quick Start

Use the npm package directly with npx — no installation required:

{
  "mcpServers": {
    "memory-mcp": {
      "command": "npx",
      "args": ["-y", "@j0hanz/memory-mcp@latest"]
    }
  }
}

[!TIP] The server uses stdio transport only; no HTTP endpoint is exposed. Stdout must not be polluted by custom logging.

Or run with Docker:

docker run --rm -i ghcr.io/j0hanz/memory-mcp:latest

Client Configuration

{
  "servers": {
    "memory-mcp": {
      "command": "npx",
      "args": ["-y", "@j0hanz/memory-mcp@latest"]
    }
  }
}

code --add-mcp '{"name":"memory-mcp","command":"npx","args":["-y","@j0hanz/memory-mcp@latest"]}'

code-insiders --add-mcp '{"name":"memory-mcp","command":"npx","args":["-y","@j0hanz/memory-mcp@latest"]}'

{
  "mcpServers": {
    "memory-mcp": {
      "command": "npx",
      "args": ["-y", "@j0hanz/memory-mcp@latest"]
    }
  }
}

{
  "mcpServers": {
    "memory-mcp": {
      "command": "npx",
      "args": ["-y", "@j0hanz/memory-mcp@latest"]
    }
  }
}

claude mcp add memory-mcp -- npx -y @j0hanz/memory-mcp@latest

{
  "mcpServers": {
    "memory-mcp": {
      "command": "npx",
      "args": ["-y", "@j0hanz/memory-mcp@latest"]
    }
  }
}

# Pull and run (stdio mode)
docker run --rm -i \
  -e MEMORY_DB_PATH=/data/memory.db \
  -v memory-data:/data \
  ghcr.io/j0hanz/memory-mcp:latest

{
  "mcpServers": {
    "memory-mcp": {
      "command": "docker",
      "args": [
        "run",
        "--rm",
        "-i",
        "-e",
        "MEMORY_DB_PATH=/data/memory.db",
        "-v",
        "memory-data:/data",
        "ghcr.io/j0hanz/memory-mcp:latest"
      ]
    }
  }
}

Documentation Maintenance

• Owner: maintainers updating MCP behavior in src/ must update README.md and affected mcp/ mirror pages in the same PR.
• Link/version policy: use pinned https://modelcontextprotocol.io/specification/2025-11-25/... links for protocol references; avoid latest and mixed legacy targets.
• Drift-check checklist:
  • Re-verify capability declarations in src/server.ts.
  • Reconcile tool/resource/prompt docs with src/tools/index.ts, src/resources/index.ts, and src/prompts/index.ts.
  • Confirm limitations/gotchas in src/instructions.md match runtime behavior.
• Validation commands: npm run type-check, npm run test:fast, npm run build.

MCP Surface

Tools Summary

store_memory

Store a new memory with content, tags, and optional type/importance. Idempotent — storing the same content+tags returns the existing hash with created: false.

Returns: { hash, created }

store_memories

Store multiple memories in one transaction (max 50 items).

Returns: { items, succeeded, failed }

get_memory

Retrieve one memory by its SHA-256 hash.

Returns: Memory or { ok: false, error } on E_NOT_FOUND.

update_memory

Update content and optionally tags for an existing memory. Returns both hashes.

Returns: { old_hash, new_hash }

delete_memory

Delete one memory by hash. Cascades to related relationship rows.

Returns: { hash, deleted }

delete_memories

Delete multiple memories by hash in one transaction.

Returns: { items, succeeded, failed }

search_memories

Full-text search over memory content and tags using FTS5. Supports importance and type filters with cursor pagination.

Returns: { memories, total_returned, nextCursor? }

create_relationship

Create a directed relationship edge between two memories. Idempotent.

Suggested relation_type values: related_to, causes, depends_on, parent_of, child_of, supersedes, contradicts, supports, references.

Returns: { created }

delete_relationship

Delete one directed relationship edge.

Returns: { deleted } or { ok: false, error } on E_NOT_FOUND.

get_relationships

Retrieve relationships for a memory, with optional direction filter.

Returns: { relationships, count }

Each relationship includes from_hash, to_hash, relation_type, created_at, linked_hash, linked_content, and linked_tags.

recall

Search memories by full-text query, then traverse the relationship graph up to depth hops via BFS. Emits MCP progress notifications per hop.

Returns: { memories, graph, depth_reached, aborted?, nextCursor? }

Each item in graph uses the shape:

{ "from_hash": "...", "to_hash": "...", "relation_type": "..." }

[!NOTE] aborted: true indicates the traversal hit a safety limit (RECALL_MAX_FRONTIER_SIZE, RECALL_MAX_EDGE_ROWS, or RECALL_MAX_VISITED_NODES). Partial results are still returned.

retrieve_context

Search memories and return relevance-ranked results that fit within a caller-specified token budget. Eliminates manual pagination and token counting for context window management.

Returns: { memories, estimated_tokens, truncated }

[!TIP] Token estimation is approximate (content length ÷ 4). truncated: true means the budget was reached before all candidates were included.

memory_stats

Return aggregate memory and relationship stats. Takes no input.

Returns:

{
  "memories": {
    "total": 0,
    "oldest": null,
    "newest": null,
    "avg_importance": null
  },
  "relationships": { "total": 0 },
  "by_type": {}
}

Resources

Prompts

Configuration

Environment Variables

[!IMPORTANT] If MEMORY_DB_PATH is relative (including the default memory_db/memory.db), it resolves from the process working directory.

[!TIP] Add memory_db/ to your .gitignore to keep the database out of version control — it contains local session data and should not be shared or committed.

Limits and Constraints

[!NOTE] Cursor values are opaque base64url-encoded tokens. Treat them as opaque and do not parse them.

Security

• Transport is stdio-only (StdioServerTransport) — no HTTP endpoints.
• Fatal process errors are written to stderr; stdout must remain clean for the MCP protocol.
• All inputs are validated with strict Zod schemas and bounded field constraints before any database access.
• Hashes are validated against a lowercase 64-char SHA-256 hex regex.
• Search input is tokenized to alphanumeric terms before FTS MATCH execution (non-alphanumeric characters act as delimiters, preventing FTS injection).
• SQLite foreign keys are enabled; relationship rows cascade-delete when a memory is removed.

Development

Install dependencies:

npm install

Core scripts:

Inspect with MCP Inspector:

npx @modelcontextprotocol/inspector node dist/index.js

Build & Release

GitHub Actions release workflow (.github/workflows/release.yml) handles versioning, validation, and publishing via a single workflow_dispatch trigger:

workflow_dispatch (patch / minor / major / custom)
    │
    ▼
  release — bump package.json + server.json → lint → type-check → test → build → tag → GitHub Release
    │
    ├──► publish-npm ──► publish-mcp   (npm Trusted Publishing OIDC → MCP Registry)
    │
    └──► publish-docker                (GHCR, linux/amd64 + linux/arm64)

Trigger a release:

gh workflow run release.yml -f bump=patch

Or use the GitHub UI: Actions → Release → Run workflow.

[!NOTE] npm publishing uses OIDC Trusted Publishing — no NPM_TOKEN secret required. MCP Registry uses GitHub OIDC. Docker uses the built-in GITHUB_TOKEN.

Troubleshooting

License

MIT