Mono Memory MCP

One shared brain for every AI on your team — persistent across sessions, searchable, always in sync.

A lightweight, self-hosted MCP server that gives your AI coding assistants long-term memory. Built for teams where multiple people use AI-powered editors (Claude Code, Cursor, Windsurf) and need their AIs to remember past decisions, share discoveries, and stay aligned — without re-explaining everything every session.

The Problem

• Your AI assistant forgets everything when a session ends.
• Each team member's AI works in isolation — no shared knowledge.
• Critical decisions, bug fixes, and architectural context get lost between sessions.

The Solution

Mono Memory gives your team's AI assistants a shared, persistent memory backed by a single SQLite file. Any AI can save and retrieve observations, project context, and decisions — across sessions, across team members.

Why "Mono"? — Like a monorepo manages all code in one place, Mono Memory manages all your team's AI knowledge in one server.

How It Works

Session 1 (Alice — morning)
├─ AI discovers a tricky bug in auth logic
├─ → memory_save: "JWT refresh token race condition fix — added mutex lock"
└─ Session ends. AI forgets everything.

Session 2 (Bob — afternoon)
├─ AI starts working on auth-related feature
├─ → memory_search: "auth"
├─ ← Gets Alice's bug fix context instantly
└─ Avoids the same pitfall, builds on her solution.

Session 3 (Alice — next day)
├─ → memory_timeline: project="my-app", since="2025-03-01"
└─ ← Sees everything the team's AIs learned this week.

Every observation is stored in a shared SQLite database. Any team member's AI can save and query it through 6 MCP tools.

Use Cases

Solo Developer

• Session continuity — Your AI remembers yesterday's debugging insights, architectural decisions, and TODO notes without you copy-pasting context.
• Project context — Store your project's architecture, conventions, and API specs once. Your AI loads them on demand instead of re-reading files every session.

Team (2-10 developers)

• Shared knowledge base — One person's AI discovers a gotcha? Everyone's AI knows about it.
• Onboarding — New team members' AIs instantly access the full history of decisions and patterns.
• Cross-project awareness — Working on the frontend? Search what the backend team's AI learned about the API yesterday.

Multi-project

• Centralized memory — One server, multiple projects. Search across all or filter by project.
• Timeline view — See the evolution of decisions across your entire organization.

Features

• 6 tools — save, get, search, timeline, init, context
• SQLite storage — zero-config, WAL mode, single-file database
• Streamable HTTP — network-ready transport for team use
• Environment variable config — host, port, database path
• Multi-project — multiple authors and projects, keyword search, timeline view

---

Quick Start

There are two roles: Host (runs the server) and Client (connects via plugin).

Host: Start the Server

The host is the person (or machine) that runs the Mono Memory server for the team.

git clone https://github.com/potato-castle/mono-memory-mcp.git
cd mono-memory-mcp
uv run python server.py

The server starts on http://0.0.0.0:8765/mcp (streamable-http). Share this URL with your team — replace 0.0.0.0 with your machine's IP address (e.g. http://192.168.0.10:8765/mcp).

Custom configuration:

# Change port
MONO_MEMORY_PORT=9000 python server.py

# Change database directory
MONO_MEMORY_DB_DIR=/path/to/data python server.py

# Run in background
nohup python server.py > /tmp/mono-memory.log 2>&1 &

Client: Install the Plugin (Claude Code)

Clients do not need to clone the repo. Just run three commands in Claude Code:

1. Register the marketplace:

/plugin marketplace add potato-castle/mono-memory-mcp

2. Install the plugin:

/plugin install mono-memory-mcp@mono-memory-mcp

When prompted for scope, select "Install for you, in this repo only (local scope)". This keeps the plugin active only in the current project.

3. Run the setup skill:

/mono-memory-mcp:setup

This will prompt you for:

1. Server URL — the host's server address (e.g. http://192.168.0.10:8765/mcp)
2. Author name — your name, used to tag memories you save

The project name is automatically detected from your current directory name.

The setup will:

• Write .mcp.json in your project root (MCP server connection)
• Append auto-recording rules to CLAUDE.md (so your AI automatically saves discoveries)

Restart Claude Code to activate.

---

Tools

memory_save — Save an observation

Store a discovery, decision, debugging insight, or any knowledge.

Parameter	Required	Description
author	Yes	Author name (e.g. "alice")
project	Yes	Project name (e.g. "my-app")
content	Yes	The content to save
tags	No	Comma-separated tags (e.g. "bug,fix,api")

memory_get — Retrieve by ID

Parameter	Required	Description
id	Yes	UUID of the observation

memory_search — Keyword search

Searches both observations and project contexts.

Parameter	Required	Description
query	Yes	Search keywords (space = AND)
author	No	Filter by author
project	No	Filter by project
limit	No	Max results (default 20)

memory_timeline — Chronological view

Parameter	Required	Description
project	No	Filter by project
author	No	Filter by author
since	No	Start date (ISO 8601, e.g. "2025-01-01")
until	No	End date (ISO 8601, e.g. "2025-01-31")
limit	No	Max results (default 50)

memory_init — Initialize/update project context

Store project information by section. Same project+section overwrites (upsert).

Parameter	Required	Description
project	Yes	Project name
section	Yes	Section name (e.g. "overview", "architecture", "api")
content	Yes	Section content
author	No	Who updated it

memory_context — Retrieve project context

Parameter	Required	Description
project	Yes	Project name
section	No	Section name (omit to list all sections)

---

Usage Examples

Example 1: Save a debugging discovery

User: "Save that the login timeout was caused by Redis connection pool exhaustion."

Tool: memory_save
  project: "auth-service"
  content: "Login timeout root cause: Redis connection pool exhaustion under load. Fix: increased pool size from 10 to 50 and added retry logic in auth/session.py"
  tags: "bug,fix,redis,performance"

Response: {"status": "saved", "id": "a1b2c3d4-...", "author": "alice", "created_at": "2025-06-15T10:30:00+09:00"}

Example 2: Search for past decisions

User: "What do we know about Redis in auth-service?"

Tool: memory_search
  query: "redis"
  project: "auth-service"

Response: {"count": 2, "results": [
  {"author": "alice", "content": "Login timeout root cause: Redis connection pool...", "source": "observation"},
  {"author": "bob", "content": "Migrated Redis from 6.x to 7.x for ACL support...", "source": "observation"}
]}

Example 3: Initialize project context

User: "Set up the architecture overview for the payments project."

Tool: memory_init
  project: "payments"
  section: "architecture"
  content: "Microservice arch. Gateway (Express) -> Payment Service (FastAPI) -> Stripe API. PostgreSQL for transactions, Redis for idempotency keys."
  author: "carol"

Response: {"status": "updated", "project": "payments", "section": "architecture", "updated_at": "2025-06-15T14:00:00+09:00"}

---

Skills

/api-docs — Generate API Documentation

Generates a Swagger-style HTML API documentation page from memories stored in the mono-memory server.

/api-docs

The skill automatically:

1. Detects your project name from the current directory
2. Searches all API-related memories (endpoints, schemas, changes)
3. Generates a self-contained api-docs.html with:

• Color-coded HTTP method badges (GET, POST, PUT, DELETE, PATCH)
• Request/Response code boxes per endpoint
• Try it panels — test APIs directly from the browser
• Path parameter & query parameter input fields (Swagger-style)
• Auth type selector (Bearer, JWT, Basic Auth, API Key)
• Send button with live response display
• Copy as curl button

Prerequisite: Save some API observations first so the skill has data to work with:

memory_save(project: "my-app", content: "GET /api/users - returns paginated user list with {page} and {limit} query params", tags: "api,endpoint")
memory_save(project: "my-app", content: "POST /api/users - creates user. Request: {name, email, role}. Response: {id, name, email, created_at}", tags: "api,endpoint")
memory_init(project: "my-app", section: "api", content: "REST API base URL: /api/v1. Auth: Bearer token required.")

---

Environment Variables

Variable	Default	Description
MONO_MEMORY_HOST	0.0.0.0	Server bind address
MONO_MEMORY_PORT	8765	Server port
MONO_MEMORY_DB_DIR	./data	Directory for the SQLite database
DEFAULT_AUTHOR	(empty)	Default author name for memory_save

---

Testing

cd mono-memory-mcp
uv run python test_server.py

The test script spawns the server with an isolated temporary database and verifies all 6 tools via streamable-http.

---

Server Management

Scripts are provided in the scripts/ directory:

./scripts/start.sh      # Start the server
./scripts/stop.sh       # Stop the server
./scripts/restart.sh    # Restart the server
./scripts/logs.sh       # Tail server logs in real-time

Database location

By default: ./data/memory.db (SQLite, WAL mode)

---

CLAUDE.md Integration

The /mono-memory-mcp:setup skill automatically appends auto-recording rules to your project's CLAUDE.md. This tells your AI assistant to:

• Automatically save bugs, decisions, and discoveries to the shared memory
• Search existing memories at the start of each session
• Write all observations in English for team consistency

For manual setup, see [CLAUDE_MD_TEMPLATE.md](CLAUDE_MD_TEMPLATE.md).

---

Privacy Policy

Mono Memory MCP is a fully self-hosted, local server.

• No data leaves your machine: All data is stored in a local SQLite file.
• No telemetry: The server does not collect, transmit, or share any usage data.
• No external network calls: The server does not make any outbound HTTP requests.
• No authentication data: The server does not handle credentials or tokens for third-party services.
• Data retention: Data persists in the SQLite database until you manually delete it.

Your memory data is entirely under your control.

---

License

MIT